tenseleyflow/documentlanguagemodel / 29a900f

Browse files

Refresh Sprint 40 closeout proofs

Authored by espadonne
SHA
29a900f3fb6e9f24b4830d7c0a9479b4f76a7a3d
Parents
fdc4063
Tree
491766b

27 changed files

StatusFile+-
M docs/cookbook/audio-training.md 1 4
A docs/cookbook/choosing-a-base.md 37 0
M docs/cookbook/multimodal-training.md 11 5
M docs/format/frontmatter.md 5 5
A docs/hardware/memory-estimates.md 38 0
M docs/hardware/vl-memory.md 23 16
M docs/index.md 5 4
M mkdocs.yml 3 0
M src/dlm/base_models/registry.py 63 59
M src/dlm/base_models/resolver.py 1 0
M src/dlm/base_models/schema.py 1 0
A src/dlm/base_models/templates/qwen3thinking.jinja 14 0
M src/dlm/export/ollama/template_registry.py 21 1
A src/dlm/export/ollama/templates/qwen3thinking.gotmpl 5 0
A tests/integration/base_models/test_13_entries_scaffold.py 45 0
M tests/integration/cli/test_registry_refresh_init.py 19 25
A tests/integration/gate/test_mixtral_gate_smoke.py 54 0
M tests/unit/base_models/test_audio_registry.py 5 5
M tests/unit/base_models/test_registry.py 6 1
M tests/unit/base_models/test_registry_2026.py 41 25
M tests/unit/base_models/test_schema.py 11 1
M tests/unit/base_models/test_vl_registry.py 41 9
A tests/unit/doc/test_v12_migrator.py 36 0
M tests/unit/export/ollama/test_template_registry.py 33 3
A tests/unit/export/test_mixtral_template.py 38 0
A tests/unit/export/test_phi4_template.py 37 0
A tests/unit/export/test_qwen3_template.py 37 0
docs/cookbook/audio-training.mdmodified
@@ -10,9 +10,6 @@ spoken-corpus workflow end-to-end: scaffold → drop clips + transcripts
1010
   24 GB VRAM. Qwen2-Audio-7B-Instruct fp16 weighs ~15 GB; the 16 GB
1111
   consumer GPUs don't fit this base without quantization (4-bit audio
1212
   training is deferred).
13
-- A Hugging Face account with the [Qwen2-Audio-7B-Instruct terms
14
-  accepted](https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct) and
15
-  `HF_TOKEN` exported.
1613
 - Qwen2-Audio cached locally (`huggingface-cli download
1714
   Qwen/Qwen2-Audio-7B-Instruct`). First train without this triggers
1815
   the download automatically.
@@ -22,7 +19,7 @@ spoken-corpus workflow end-to-end: scaffold → drop clips + transcripts
2219
 ## Step 1 — Scaffold an audio `.dlm`
2320
2421
 ```bash
25
-dlm init my-audio.dlm --audio --i-accept-license
22
+dlm init my-audio.dlm --audio
2623
 ```
2724
2825
 `--audio` pins the base to `qwen2-audio-7b-instruct` and emits a
docs/cookbook/choosing-a-base.mdadded
@@ -0,0 +1,37 @@
1
+# Choosing a base
2
+
3
+The fastest way to pick a DLM base is to decide three things first:
4
+
5
+1. Do you need plain text, multimodal vision, or audio?
6
+2. Do you want the most permissive license possible, or are gated rows fine?
7
+3. Are you targeting Apple Silicon, a mid-size CUDA card, or a large CUDA box?
8
+
9
+## Quick picks
10
+
11
+| If you want… | Start with… | Why |
12
+|---|---|---|
13
+| Fast local iteration on almost any laptop | `smollm2-135m` | Tiny, cheap, and ideal for testing authoring loops. |
14
+| Best general-purpose 2026 text base around the 4B tier | `qwen3-4b` | Strong default quality, permissive license, and current-generation tokenizer/chat behavior. |
15
+| A reasoning-first 1.7B profile | `qwen3-1.7b-thinking` | Same upstream Qwen3 weights, but a curated reasoning-profile key with cooler defaults. |
16
+| Fully open-model story | `olmo-2-7b-instruct` | Open weights and open-data lineage make it the cleanest reproducibility pitch. |
17
+| Apache sparse-MoE experiments | `mixtral-8x7b-instruct` | First `text-moe` row in the registry; pairs with the learned gate work. |
18
+| Small gated text base | `gemma-2-2b-it` | Useful when Gemma’s instruction style or ecosystem matters more than license friction. |
19
+| Larger gated text base | `gemma-2-9b-it` | Upper-tier Gemma pick; large enough to want real GPU planning. |
20
+| Large multimodal capability | `mistral-small-3.1-24b-instruct` | Strongest shipped VL row, but large-CUDA-first. |
21
+| Safe default multimodal row on a smaller box | `qwen2-vl-2b-instruct` | Permissive, solid, and compatible with the current generic VL runtime. |
22
+| Audio-language training | `qwen2-audio-7b-instruct` | Current shipped audio row; open-license and no longer gated on HF. |
23
+
24
+## Notes on the sharp edges
25
+
26
+- `llama-3.3-8b-instruct` is still treated like the Llama family in DLM’s policy surface: acceptance required, not redistributable, and intended for users who already know they want the Llama line.
27
+- `internvl2-2b` and `internvl3-2b` are registry-visible planning targets, but the current generic VL runtime still refuses the InternVL family until DLM owns its custom processor/collator contract.
28
+- `mistral-small-3.1-24b-instruct` is intentionally refused on MPS by default. It is a real shipped row, just not a casual laptop target.
29
+
30
+## Hardware-first view
31
+
32
+- Apple Silicon, 16 GB: `smollm2-*`, `qwen2.5-*`, `qwen3-1.7b`, and `qwen3-4b` are the comfortable text picks; `qwen2-vl-2b-instruct` is the safer VL row.
33
+- Apple Silicon, 32 GB+: `qwen3-8b`, `gemma-2-2b-it`, and `phi-4-mini-reasoning` become practical. Large VL rows still need caution.
34
+- CUDA, 24 GB: this is where `gemma-2-9b-it`, `mixtral-8x7b-instruct`, and the heavier multimodal rows start becoming realistic.
35
+- CUDA, 48 GB+: this is the intended home for `mistral-small-3.1-24b-instruct`.
36
+
37
+See [hardware/memory-estimates](../hardware/memory-estimates.md) for the text-family budget table and [hardware/vl-memory](../hardware/vl-memory.md) for the VL rows.
docs/cookbook/multimodal-training.mdmodified
@@ -28,7 +28,7 @@ drop real images into that path before the first train).
2828
2929
 ### Picking a different VL base
3030
31
-Four VL bases ship in the registry today:
31
+Five VL bases ship in the registry today:
3232
3333
 ```bash
3434
 # Permissive + Apache-2.0 + strong general-purpose VL (pinned 672²):
@@ -37,6 +37,9 @@ dlm init my-diagrams.dlm --multimodal --base qwen2-vl-2b-instruct
3737
 # MIT-licensed, smallest per-image footprint (448²):
3838
 dlm init my-diagrams.dlm --multimodal --base internvl2-2b
3939
40
+# Newer InternVL planning row (dynamic 448-tiling, still runtime-deferred):
41
+dlm init my-diagrams.dlm --multimodal --base internvl3-2b
42
+
4043
 # Largest-capability VL row, CUDA-first (pinned 1540²):
4144
 dlm init my-diagrams.dlm --multimodal --base mistral-small-3.1-24b-instruct
4245
@@ -50,8 +53,10 @@ base-selection matrix. **Heads-up on InternVL2**: the row is visible in
5053
 the registry, but on the current stack DLM now refuses it for actual
5154
 prompt/train/HF-snapshot-export work. The upstream family still needs a
5255
 custom processor/collator path for its tokenizer-only `AutoProcessor`,
53
-`<image>` expansion, and `image_flags` forward contract. That same
54
-family gap is the reason `internvl3-2b` has not been added yet.
56
+`<image>` expansion, and `image_flags` forward contract. The same
57
+family gap applies to `internvl3-2b` as well: it is now registry-
58
+visible and scaffoldable, but the generic runtime still refuses the
59
+whole InternVL family until DLM owns that custom contract.
5560
 **Heads-up on Mistral Small 3.1**: it is a real VL registry row now,
5661
 but it is intentionally treated as a large-CUDA-first base. `dlm
5762
 doctor` refuses it on Apple Silicon by default unless you explicitly
@@ -147,8 +152,9 @@ coverage of the base's arch class and routes to one of three paths:
147152
   None of the registered bases hit this verdict at the pinned tag.
148153
 - **UNSUPPORTED** — llama.cpp doesn't know the arch at all. Falls
149154
   back to HF-snapshot with an actionable banner naming the arch
150
-  class and the vendored tag. **paligemma-3b-mix-224** and
151
-  **internvl2-2b** are UNSUPPORTED at the pinned tag.
155
+  class and the vendored tag. **paligemma-3b-mix-224**,
156
+  **internvl2-2b**, and **internvl3-2b** are UNSUPPORTED at the
157
+  pinned tag.
152158
153159
 See [docs/hardware/vl-memory.md](../hardware/vl-memory.md#llamacpp-gguf-support-matrix-sprint-354)
154160
 for the current support verdicts; bump the vendored tag with
docs/format/frontmatter.mdmodified
@@ -35,12 +35,12 @@ it; don't edit it by hand.
3535
 The shipped registry is broader than this quick-start table. Current
3636
 additions include:
3737
38
-- 2026 text-family refresh rows: `qwen3-1.7b`, `qwen3-4b`, `qwen3-8b`,
39
-  `llama-3.3-8b-instruct`, `phi-4-mini-reasoning`, `gemma-2-2b-it`,
40
-  `gemma-2-9b-it`, `smollm3-3b`, `olmo-2-7b-instruct`, and
41
-  `mixtral-8x7b-instruct`.
38
+- 2026 text-family refresh rows: `qwen3-1.7b`, `qwen3-1.7b-thinking`,
39
+  `qwen3-4b`, `qwen3-8b`, `llama-3.3-8b-instruct`,
40
+  `phi-4-mini-reasoning`, `gemma-2-2b-it`, `gemma-2-9b-it`,
41
+  `smollm3-3b`, `olmo-2-7b-instruct`, and `mixtral-8x7b-instruct`.
4242
 - Vision-language rows: `paligemma-3b-mix-224`,
43
-  `qwen2-vl-2b-instruct`, `internvl2-2b`, and
43
+  `qwen2-vl-2b-instruct`, `internvl2-2b`, `internvl3-2b`, and
4444
   `mistral-small-3.1-24b-instruct`.
4545
 - Audio-language row: `qwen2-audio-7b-instruct`.
4646
docs/hardware/memory-estimates.mdadded
@@ -0,0 +1,38 @@
1
+# Memory estimates
2
+
3
+These are planning numbers, not a promise. DLM’s doctor still does the
4
+real refusal/fit decision, but the table below is the quick mental map
5
+for the Sprint 40 refresh rows that changed the most user expectations.
6
+
7
+## Text-family checkpoints
8
+
9
+| Base | fp16 weights | Practical target |
10
+|---|---:|---|
11
+| `qwen3-8b` | ~16 GB | 24 GB CUDA or high-memory Apple Silicon for LoRA; lighter inference on smaller boxes. |
12
+| `llama-3.3-8b-instruct` | ~16.5 GB | Same class as other 8B text rows: real GPU planning required for training. |
13
+| `gemma-2-9b-it` | ~18 GB | 24 GB CUDA is the comfortable floor. |
14
+| `mistral-small-3.1-24b-instruct` | ~48 GB | Large-CUDA-first. Refused on MPS by default unless forced. |
15
+
16
+## What the doctor is approximating
17
+
18
+For LoRA/QLoRA, the planner estimates:
19
+
20
+- base weights at the chosen load precision
21
+- activation memory from `sequence_len × micro_batch × layers`
22
+- optimizer state for the trainable adapter params
23
+- LoRA parameter storage
24
+- a 20% safety margin on top
25
+
26
+That estimator lives in `src/dlm/hardware/memory.py` and is intentionally conservative.
27
+
28
+## Rules of thumb
29
+
30
+- 8B-class rows are where laptop experimentation starts turning into real hardware planning.
31
+- 9B-class rows are usually fine on 24 GB CUDA, but not “casual” on smaller hosts.
32
+- 24B-class rows are not broad consumer defaults. In DLM they are treated as explicit high-capacity picks.
33
+- MPS can be surprisingly good for text LoRA, but DLM now refuses oversized bases like `mistral-small-3.1-24b-instruct` by default because unified memory headroom disappears too quickly.
34
+
35
+## Related
36
+
37
+- [Choosing a base](../cookbook/choosing-a-base.md)
38
+- [Vision-language memory budget](vl-memory.md)
docs/hardware/vl-memory.mdmodified
@@ -1,7 +1,7 @@
11
 # Vision-language memory budget
22
3
-Four VL rows now ship in the registry: **PaliGemma-3B-mix-224**,
4
-**Qwen2-VL-2B-Instruct**, **InternVL2-2B**, and
3
+Five VL rows now ship in the registry: **PaliGemma-3B-mix-224**,
4
+**Qwen2-VL-2B-Instruct**, **InternVL2-2B**, **InternVL3-2B**, and
55
 **Mistral-Small-3.1-24B-Instruct-2503**. Each row carries a pinned
66
 preprocessing plan; dynamic-resolution support (Qwen2-VL's native
77
 capability, Mistral Small 3.1's longer-edge policy, and the broader
@@ -10,9 +10,10 @@ work so the current `VlPreprocessorPlan` cache key stays stable.
1010
1111
 **Reality check.** The generic VL train/prompt path is complete today
1212
 for PaliGemma, Qwen2-VL, and Mistral Small 3.1. InternVL2 remains
13
-registry-visible for planning and future support, but on the current
14
-transformers stack its HF path still exposes a tokenizer-only
15
-`AutoProcessor` and needs a custom collator/runtime contract. DLM now
13
+registry-visible for planning and future support, and InternVL3 now
14
+joins it under the same honest caveat: on the current transformers
15
+stack the InternVL family still exposes a tokenizer-only
16
+`AutoProcessor` and needs a custom collator/runtime contract. DLM
1617
 refuses that family with a clear error instead of pretending the
1718
 generic VL path is enough.
1819
@@ -23,6 +24,7 @@ generic VL path is enough.
2324
 | paligemma-3b-mix-224      | Gemma (gated) | The cleanest PEFT path + proven chart/doc QA; accept the Gemma license first. |
2425
 | qwen2-vl-2b-instruct      | Apache-2.0 | Permissive licensing + strong general-purpose VL; dynamic-res is capped to 672² in v1 but native runtime supports more. |
2526
 | internvl2-2b              | MIT        | Registry-visible planning target for a future custom InternVL path; current train/prompt/export-snapshot flows refuse it on this stack. |
27
+| internvl3-2b              | Apache-2.0 | Newer InternVL planning target with dynamic 448-tiling and `trust_remote_code`; currently registry-visible but still refused by the generic runtime. |
2628
 | mistral-small-3.1-24b-instruct | Apache-2.0 | Highest-capability VL row in the registry today; targets large CUDA boxes first and is refused on MPS by default unless you explicitly force it. |
2729
2830
 ## PaliGemma-3B-mix-224 (224×224, fp16)
@@ -70,11 +72,14 @@ between vision + text tokens. Gradient checkpointing on the tower
7072
 trims ~30% of peak; `training.gradient_checkpointing: true` in
7173
 frontmatter enables it.
7274
73
-## InternVL2-2B (448×448, fp16)
75
+## InternVL2-2B / InternVL3-2B (448×448, fp16)
7476
7577
 InternVL2 uses ViT-L/14 + pixel-shuffle 2×2 so 448² input yields 256
7678
 image tokens per 448-tile — the smallest InternVL-family budget and
77
-the cheapest of the four rows on paper.
79
+the cheapest of the registry rows on paper. InternVL3 keeps the same
80
+448 target size but switches the registry row to `resize_policy:
81
+dynamic` and a user-visible `<image>` placeholder while still
82
+expanding into the same hidden InternVL context window at runtime.
7883
7984
 | Config          | Base weights | Adapter | Activations | Total (peak) |
8085
 |-----------------|-------------:|--------:|------------:|-------------:|
@@ -86,15 +91,15 @@ the cheapest of the four rows on paper.
8691
 memory alone. 12 GB CUDA would handle batch=1; 16 GB CUDA would handle
8792
 batch=4.
8893
89
-**Current runtime status.** This row is not trainable/promptable via
90
-the generic VL path today. InternVL2 ships as `InternVLChatModel`, a
91
-custom remote-code family whose upstream runtime expands `<image>` into
92
-repeated `<IMG_CONTEXT>` spans and threads `image_flags` through the
93
-forward pass. On the current stack, `AutoProcessor.from_pretrained(...)`
94
-resolves to a tokenizer-only object, so DLM refuses the family early
95
-instead of failing later inside the model. Keep the budget numbers here
96
-for planning, but use PaliGemma, Qwen2-VL, or Mistral Small 3.1 for
97
-actual runs today.
94
+**Current runtime status.** These rows are not trainable/promptable via
95
+the generic VL path today. InternVL2 and InternVL3 both ship as
96
+`InternVLChatModel`, a custom remote-code family whose upstream runtime
97
+expands `<image>` into repeated `<IMG_CONTEXT>` spans and threads
98
+`image_flags` through the forward pass. On the current stack,
99
+`AutoProcessor.from_pretrained(...)` resolves to a tokenizer-only
100
+object, so DLM refuses the family early instead of failing later inside
101
+the model. Keep the budget numbers here for planning, but use
102
+PaliGemma, Qwen2-VL, or Mistral Small 3.1 for actual runs today.
98103
99104
 ## Mistral Small 3.1 24B Instruct (pinned 1540×1540, fp16)
100105
@@ -129,6 +134,7 @@ by `scripts/bump-llama-cpp.sh bump <tag>`):
129134
 | paligemma-3b-mix-224      | PaliGemmaForConditionalGeneration   | UNSUPPORTED  |
130135
 | qwen2-vl-2b-instruct      | Qwen2VLForConditionalGeneration     | SUPPORTED    |
131136
 | internvl2-2b              | InternVLChatModel                   | UNSUPPORTED  |
137
+| internvl3-2b              | InternVLChatModel                   | UNSUPPORTED  |
132138
133139
 **UNSUPPORTED** means `dlm export` falls back to the HF-snapshot path
134140
 with an actionable banner. **SUPPORTED** means single-file VL GGUF
@@ -175,6 +181,7 @@ with the preprocessing plan:
175181
 |---------------------------|------------:|----------------:|
176182
 | paligemma-3b-mix-224      |     224×224 |        ~0.5 MB  |
177183
 | internvl2-2b              |     448×448 |        ~2.0 MB  |
184
+| internvl3-2b              |     448×448 |        ~2.0 MB  |
178185
 | qwen2-vl-2b-instruct      |     672×672 |        ~4.5 MB  |
179186
 | mistral-small-3.1-24b-instruct | 1540×1540 |       ~23.5 MB  |
180187
docs/index.mdmodified
@@ -19,10 +19,11 @@ you to run a 70B model you can't afford. DLM sits in the gap:
1919
   control is both the prose you're training on and the configuration
2020
   for how the training runs. Edit, retrain, share.
2121
 - **Real pretrained bases.** SmolLM2-135M for fast iteration; newer
22
-  registry rows like Qwen3, Llama 3.3, Gemma 2, SmolLM3, Phi-4-mini-
23
-  reasoning, OLMo-2, Mixtral, and Mistral Small 3.1 cover current
24
-  text, sparse-MoE, and multimodal use cases. No from-scratch
25
-  transformers, no toy experiments.
22
+  registry rows like Qwen3 (including a reasoning-profile key),
23
+  Llama 3.3, Gemma 2, SmolLM3, Phi-4-mini-reasoning, OLMo-2, Mixtral,
24
+  Mistral Small 3.1, and InternVL3 cover current text, sparse-MoE,
25
+  and multimodal planning use cases. No from-scratch transformers,
26
+  no toy experiments.
2627
 - **Deterministic by contract.** Same document + same hardware tier +
2728
   pinned versions produce bit-identical adapters. [Determinism](determinism.md)
2829
   is a first-class feature.
mkdocs.ymlmodified
@@ -62,6 +62,7 @@ nav:
6262
       - .dlm/ignore: format/dlm-ignore.md
6363
   - CLI reference: cli/reference.md
6464
   - Cookbook:
65
+      - Choosing a base: cookbook/choosing-a-base.md
6566
       - Coding tutor: cookbook/coding-tutor.md
6667
       - Domain knowledge base: cookbook/domain-kb.md
6768
       - Writing partner: cookbook/writing-partner.md
@@ -82,5 +83,7 @@ nav:
8283
   - Architecture: architecture.md
8384
   - Determinism: determinism.md
8485
   - Hardware:
86
+      - Memory estimates: hardware/memory-estimates.md
87
+      - Vision-language memory: hardware/vl-memory.md
8588
       - AMD ROCm: hardware/rocm.md
8689
   - Troubleshooting: troubleshooting.md
src/dlm/base_models/registry.pymodified
@@ -107,10 +107,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
107107
     BaseModelSpec(
108108
         key="qwen3-1.7b",
109109
         hf_id="Qwen/Qwen3-1.7B",
110
-        # Placeholder SHA: format-valid, not a real HF commit. The
111
-        # weekly `scripts/refresh-registry.py --check` run surfaces
112
-        # drift and prints the live value for manual review.
113
-        revision="1a2b3c4d5e6f7890abcdeffedcba0987654321ab",
110
+        revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
114111
         architecture="Qwen3ForCausalLM",
115112
         params=1_700_000_000,
116113
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -126,10 +123,29 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
126123
         recommended_seq_len=2048,
127124
         reasoning_tuned=True,
128125
     ),
126
+    BaseModelSpec(
127
+        key="qwen3-1.7b-thinking",
128
+        hf_id="Qwen/Qwen3-1.7B",
129
+        revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
130
+        architecture="Qwen3ForCausalLM",
131
+        params=1_700_000_000,
132
+        target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
133
+        template="qwen3thinking",
134
+        gguf_arch="qwen3",
135
+        tokenizer_pre="qwen2",
136
+        license_spdx="Apache-2.0",
137
+        license_url="https://huggingface.co/Qwen/Qwen3-1.7B/blob/main/LICENSE",
138
+        requires_acceptance=False,
139
+        redistributable=True,
140
+        size_gb_fp16=3.4,
141
+        context_length=32_768,
142
+        recommended_seq_len=2048,
143
+        reasoning_tuned=True,
144
+    ),
129145
     BaseModelSpec(
130146
         key="qwen3-4b",
131147
         hf_id="Qwen/Qwen3-4B",
132
-        revision="2b3c4d5e6f7890abcdeffedcba0987654321abc2",
148
+        revision="1cfa9a7208912126459214e8b04321603b3df60c",
133149
         architecture="Qwen3ForCausalLM",
134150
         params=4_000_000_000,
135151
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -148,7 +164,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
148164
     BaseModelSpec(
149165
         key="qwen3-8b",
150166
         hf_id="Qwen/Qwen3-8B",
151
-        revision="3c4d5e6f7890abcdeffedcba0987654321abc2d3",
167
+        revision="b968826d9c46dd6066d109eabc6255188de91218",
152168
         architecture="Qwen3ForCausalLM",
153169
         params=8_000_000_000,
154170
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -225,10 +241,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
225241
     BaseModelSpec(
226242
         key="smollm3-3b",
227243
         hf_id="HuggingFaceTB/SmolLM3-3B",
228
-        # Placeholder SHA: format-valid, not a real HF commit. The
229
-        # weekly `scripts/refresh-registry.py --check` run surfaces
230
-        # drift and prints the live value for manual review.
231
-        revision="5e6f7890abcdeffedcba0987654321abc2d3e4f5",
244
+        revision="a07cc9a04f16550a088caea529712d1d335b0ac1",
232245
         architecture="SmolLM3ForCausalLM",
233246
         params=3_000_000_000,
234247
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -247,10 +260,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
247260
     BaseModelSpec(
248261
         key="olmo-2-7b-instruct",
249262
         hf_id="allenai/OLMo-2-1124-7B-Instruct",
250
-        # Placeholder SHA: format-valid, not a real HF commit. The
251
-        # weekly `scripts/refresh-registry.py --check` run surfaces
252
-        # drift and prints the live value for manual review.
253
-        revision="6f7890abcdeffedcba0987654321abc2d3e4f5a6",
263
+        revision="470b1fba1ae01581f270116362ee4aa1b97f4c84",
254264
         architecture="Olmo2ForCausalLM",
255265
         params=7_000_000_000,
256266
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -268,10 +278,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
268278
     BaseModelSpec(
269279
         key="gemma-2-2b-it",
270280
         hf_id="google/gemma-2-2b-it",
271
-        # Placeholder SHA: format-valid, not a real HF commit. The
272
-        # weekly `scripts/refresh-registry.py --check` run surfaces
273
-        # drift and prints the live value for manual review.
274
-        revision="7a890abcdeffedcba0987654321abc2d3e4f5a6b",
281
+        revision="299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8",
275282
         architecture="Gemma2ForCausalLM",
276283
         params=2_600_000_000,
277284
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -289,10 +296,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
289296
     BaseModelSpec(
290297
         key="gemma-2-9b-it",
291298
         hf_id="google/gemma-2-9b-it",
292
-        # Placeholder SHA: format-valid, not a real HF commit. The
293
-        # weekly `scripts/refresh-registry.py --check` run surfaces
294
-        # drift and prints the live value for manual review.
295
-        revision="8f90abcdeffedcba0987654321abc2d3e4f5a6b7",
299
+        revision="11c9b309abf73637e4b6f9a3fa1e92e615547819",
296300
         architecture="Gemma2ForCausalLM",
297301
         params=9_000_000_000,
298302
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -382,10 +386,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
382386
     BaseModelSpec(
383387
         key="phi-4-mini-reasoning",
384388
         hf_id="microsoft/Phi-4-mini-reasoning",
385
-        # Placeholder SHA: format-valid, not a real HF commit. The
386
-        # weekly `scripts/refresh-registry.py --check` run surfaces
387
-        # drift and prints the live value for manual review.
388
-        revision="9a0bcdeffedcba0987654321abc2d3e4f5a6b7c8",
389
+        revision="0e3b1e2d02ee478a3743abe3f629e9c0cb722e0a",
389390
         architecture="Phi3ForCausalLM",
390391
         params=3_800_000_000,
391392
         target_modules=["qkv_proj", "o_proj", "gate_up_proj", "down_proj"],
@@ -411,10 +412,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
411412
     BaseModelSpec(
412413
         key="mixtral-8x7b-instruct",
413414
         hf_id="mistralai/Mixtral-8x7B-Instruct-v0.1",
414
-        # Placeholder SHA: format-valid, not a real HF commit. The
415
-        # weekly `scripts/refresh-registry.py --check` run surfaces
416
-        # drift and prints the live value for manual review.
417
-        revision="bc0deffedcba0987654321abc2d3e4f5a6b7c8d9",
415
+        revision="eba92302a2861cdc0098cc54bc9f17cb2c47eb61",
418416
         architecture="MixtralForCausalLM",
419417
         params=46_700_000_000,
420418
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -443,10 +441,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
443441
     BaseModelSpec(
444442
         key="mistral-small-3.1-24b-instruct",
445443
         hf_id="mistralai/Mistral-Small-3.1-24B-Instruct-2503",
446
-        # Placeholder SHA: format-valid, not a real HF commit. The
447
-        # weekly `scripts/refresh-registry.py --check` run surfaces
448
-        # drift and prints the live value for manual review.
449
-        revision="ab0cdeffedcba0987654321abc2d3e4f5a6b7c8d",
444
+        revision="68faf511d618ef198fef186659617cfd2eb8e33a",
450445
         architecture="Mistral3ForConditionalGeneration",
451446
         params=24_000_000_000,
452447
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -484,14 +479,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
484479
     BaseModelSpec(
485480
         key="paligemma-3b-mix-224",
486481
         hf_id="google/paligemma-3b-mix-224",
487
-        # Placeholder SHA: format-valid, not a real HF commit. The
488
-        # weekly `scripts/refresh-registry.py --check` run surfaces
489
-        # it as drift; a maintainer pastes in the observed SHA from
490
-        # the script's output. Offline probe tests skip cleanly
491
-        # until then (see tests/unit/base_models/test_vl_registry.py).
492
-        # To verify, run:
493
-        #     uv run python scripts/refresh-registry.py --check
494
-        revision="8d2f7bc9c15d71a00c14f9eb7e4c7b99c79e0a11",
482
+        revision="d1d8734c9c3ad0ccfeea4afc270faa356c2ba515",
495483
         architecture="PaliGemmaForConditionalGeneration",
496484
         params=2_900_000_000,
497485
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -529,10 +517,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
529517
     BaseModelSpec(
530518
         key="qwen2-vl-2b-instruct",
531519
         hf_id="Qwen/Qwen2-VL-2B-Instruct",
532
-        # Placeholder SHA (format-valid, not a real commit). See the
533
-        # paligemma entry for the self-healing workflow via
534
-        # `scripts/refresh-registry.py --check`.
535
-        revision="c0d1e2f3a4b5c6d7e8f9a0b1c2d3e4f5a6b7c8d9",
520
+        revision="895c3a49bc3fa70a340399125c650a463535e71c",
536521
         architecture="Qwen2VLForConditionalGeneration",
537522
         params=2_200_000_000,
538523
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -570,8 +555,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
570555
     BaseModelSpec(
571556
         key="internvl2-2b",
572557
         hf_id="OpenGVLab/InternVL2-2B",
573
-        # Placeholder SHA (format-valid, not a real commit).
574
-        revision="d1e2f3a4b5c6d7e8f9a0b1c2d3e4f5a6b7c8d9e0",
558
+        revision="e4f6747bd20f139e637642c6a058c6bd00b36919",
575559
         architecture="InternVLChatModel",
576560
         params=2_200_000_000,
577561
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -594,14 +578,37 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
594578
             num_image_tokens=256,
595579
         ),
596580
     ),
581
+    BaseModelSpec(
582
+        key="internvl3-2b",
583
+        hf_id="OpenGVLab/InternVL3-2B",
584
+        revision="899155015275a9b7338c7f4677e19c784e0e5a21",
585
+        architecture="InternVLChatModel",
586
+        params=2_000_000_000,
587
+        target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
588
+        template="internvl2",
589
+        gguf_arch="internvl3",
590
+        tokenizer_pre="internvl3",
591
+        license_spdx="Apache-2.0",
592
+        license_url="https://huggingface.co/OpenGVLab/InternVL3-2B",
593
+        requires_acceptance=False,
594
+        redistributable=True,
595
+        trust_remote_code=True,
596
+        size_gb_fp16=4.0,
597
+        context_length=32_768,
598
+        recommended_seq_len=2048,
599
+        modality="vision-language",
600
+        vl_preprocessor_plan=VlPreprocessorPlan(
601
+            target_size=(448, 448),
602
+            resize_policy="dynamic",
603
+            image_token="<image>",
604
+            num_image_tokens=256,
605
+        ),
606
+    ),
597607
     # --- Audio-language bases -----------------------------------------------
598608
     # Qwen2-Audio-7B-Instruct — Alibaba's open audio-text model. Uses
599609
     # the Qwen2 LLM backbone + a dedicated audio encoder. Apache-2.0
600
-    # but the 7B checkpoint is gated on HF via license acceptance, so
601
-    # `requires_acceptance=True` flows through the same pattern the
602
-    # Llama-3.2 / PaliGemma entries use. Redistributable under
603
-    # Apache-2.0, but not-bundled-by-default because the pack size
604
-    # (~14 GB fp16) dominates the tarball.
610
+    # and currently ungated on HF, so the registry keeps it open and
611
+    # redistributable like the other permissive Qwen rows.
605612
     #
606613
     # The 16 kHz pin + 30 s max-length match the training-time
607614
     # defaults documented in the Qwen2-Audio card. Resampling support
@@ -614,10 +621,7 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
614621
     BaseModelSpec(
615622
         key="qwen2-audio-7b-instruct",
616623
         hf_id="Qwen/Qwen2-Audio-7B-Instruct",
617
-        # Placeholder SHA (format-valid, not a real commit). See the
618
-        # paligemma entry for the self-healing workflow via
619
-        # `scripts/refresh-registry.py --check`.
620
-        revision="a1b2c3d4e5f678901234567890abcdef01234567",
624
+        revision="0a095220c30b7b31434169c3086508ef3ea5bf0a",
621625
         architecture="Qwen2AudioForConditionalGeneration",
622626
         params=8_400_000_000,
623627
         target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
@@ -626,8 +630,8 @@ _ENTRIES: tuple[BaseModelSpec, ...] = (
626630
         tokenizer_pre="qwen2",
627631
         license_spdx="Apache-2.0",
628632
         license_url="https://huggingface.co/Qwen/Qwen2-Audio-7B-Instruct",
629
-        requires_acceptance=True,
630
-        redistributable=False,
633
+        requires_acceptance=False,
634
+        redistributable=True,
631635
         size_gb_fp16=15.5,
632636
         context_length=8_192,
633637
         recommended_seq_len=2048,
src/dlm/base_models/resolver.pymodified
@@ -30,6 +30,7 @@ from dlm.base_models.schema import BaseModelSpec
3030
3131
 TemplateDialect = Literal[
3232
     "chatml",
33
+    "qwen3thinking",
3334
     "gemma2",
3435
     "smollm3",
3536
     "olmo2",
src/dlm/base_models/schema.pymodified
@@ -101,6 +101,7 @@ class BaseModelSpec(BaseModel):
101101
     target_modules: list[str] = Field(..., min_length=1)
102102
     template: Literal[
103103
         "chatml",
104
+        "qwen3thinking",
104105
         "gemma2",
105106
         "smollm3",
106107
         "olmo2",
src/dlm/base_models/templates/qwen3thinking.jinjaadded
@@ -0,0 +1,14 @@
1
+{#
2
+Qwen3 reasoning-profile reference template.
3
+
4
+Upstream keeps ChatML framing for request construction; the profile
5
+delta is the reasoning-first sampling/runtime behavior rather than a
6
+different turn wrapper.
7
+#}
8
+{%- for message in messages -%}
9
+<|im_start|>{{ message['role'] }}
10
+{{ message['content'] }}<|im_end|>
11
+{% endfor -%}
12
+{%- if add_generation_prompt -%}
13
+<|im_start|>assistant
14
+{%- endif -%}
src/dlm/export/ollama/template_registry.pymodified
@@ -23,7 +23,17 @@ from typing import Final, Literal
2323
2424
 from dlm.export.ollama.errors import TemplateRegistryError
2525
26
-Dialect = Literal["chatml", "gemma2", "smollm3", "olmo2", "llama3", "phi3", "phi4mini", "mistral"]
26
+Dialect = Literal[
27
+    "chatml",
28
+    "qwen3thinking",
29
+    "gemma2",
30
+    "smollm3",
31
+    "olmo2",
32
+    "llama3",
33
+    "phi3",
34
+    "phi4mini",
35
+    "mistral",
36
+]
2737
2838
 _TEMPLATES_DIR: Final[Path] = Path(__file__).resolve().parent / "templates"
2939
@@ -62,6 +72,16 @@ _REGISTRY: Final[dict[Dialect, DialectTemplate]] = {
6272
         # synthesize a new turn instead of yielding.
6373
         default_stops=("<|im_end|>", "<|endoftext|>", "<|im_start|>"),
6474
     ),
75
+    "qwen3thinking": DialectTemplate(
76
+        dialect="qwen3thinking",
77
+        template_path=_TEMPLATES_DIR / "qwen3thinking.gotmpl",
78
+        # Qwen3's reasoning profile still uses ChatML turn framing, but
79
+        # the upstream defaults run slightly broader sampling than the
80
+        # legacy ChatML family.
81
+        default_stops=("<|im_end|>", "<|endoftext|>", "<|im_start|>"),
82
+        default_temperature=0.6,
83
+        default_top_p=0.95,
84
+    ),
6585
     "gemma2": DialectTemplate(
6686
         dialect="gemma2",
6787
         template_path=_TEMPLATES_DIR / "gemma2.gotmpl",
src/dlm/export/ollama/templates/qwen3thinking.gotmpladded
@@ -0,0 +1,5 @@
1
+{{- if .System }}<|im_start|>system
2
+{{ .System }}<|im_end|>
3
+{{ end }}{{- range .Messages }}<|im_start|>{{ .Role }}
4
+{{ .Content }}<|im_end|>
5
+{{ end }}<|im_start|>assistant
tests/integration/base_models/test_13_entries_scaffold.pyadded
@@ -0,0 +1,45 @@
1
+"""Sprint 40 closeout mirror for the named 13-entry scaffold deliverable."""
2
+
3
+from __future__ import annotations
4
+
5
+from pathlib import Path
6
+
7
+import pytest
8
+from tests.integration.cli.test_registry_refresh_init import SPRINT40_INIT_CASES
9
+from typer.testing import CliRunner
10
+
11
+from dlm.cli.app import app
12
+from dlm.doc.parser import parse_file
13
+from dlm.doc.sections import SectionType
14
+
15
+
16
+@pytest.mark.parametrize(("base_key", "extra_flags", "expect_image_section"), SPRINT40_INIT_CASES)
17
+def test_init_scaffolds_for_all_thirteen_registry_refresh_entries(
18
+    tmp_path: Path,
19
+    base_key: str,
20
+    extra_flags: list[str],
21
+    expect_image_section: bool,
22
+) -> None:
23
+    runner = CliRunner()
24
+    home = tmp_path / "home"
25
+    doc = tmp_path / f"{base_key}.dlm"
26
+
27
+    result = runner.invoke(
28
+        app,
29
+        [
30
+            "--home",
31
+            str(home),
32
+            "init",
33
+            str(doc),
34
+            "--base",
35
+            base_key,
36
+            *extra_flags,
37
+        ],
38
+    )
39
+    assert result.exit_code == 0, result.output
40
+    parsed = parse_file(doc)
41
+    section_types = {section.type for section in parsed.sections}
42
+    if expect_image_section:
43
+        assert SectionType.IMAGE in section_types
44
+    else:
45
+        assert SectionType.IMAGE not in section_types
tests/integration/cli/test_registry_refresh_init.pymodified
@@ -1,11 +1,4 @@
1
-"""Scaffold coverage for the Sprint 40 registry refresh entries we ship.
2
-
3
-This is intentionally scoped to entries that currently exist in the
4
-registry. Two rows in the original sprint draft still need upstream
5
-reality work (`qwen3-1.7b-thinking`, `internvl3-2b`), so this test
6
-guards the refresh surface we have actually landed rather than baking
7
-stale assumptions into CI.
8
-"""
1
+"""Scaffold coverage for every Sprint 40 registry-refresh entry."""
92
103
 from __future__ import annotations
114
@@ -18,24 +11,25 @@ from dlm.cli.app import app
1811
 from dlm.doc.parser import parse_file
1912
 from dlm.doc.sections import SectionType
2013
21
-
22
-@pytest.mark.parametrize(
23
-    ("base_key", "extra_flags", "expect_image_section"),
24
-    [
25
-        ("qwen3-1.7b", [], False),
26
-        ("qwen3-4b", [], False),
27
-        ("qwen3-8b", [], False),
28
-        ("llama-3.3-8b-instruct", ["--i-accept-license"], False),
29
-        ("phi-4-mini-reasoning", [], False),
30
-        ("gemma-2-2b-it", ["--i-accept-license"], False),
31
-        ("gemma-2-9b-it", ["--i-accept-license"], False),
32
-        ("mistral-small-3.1-24b-instruct", ["--multimodal"], True),
33
-        ("smollm3-3b", [], False),
34
-        ("olmo-2-7b-instruct", [], False),
35
-        ("mixtral-8x7b-instruct", [], False),
36
-    ],
14
+SPRINT40_INIT_CASES: tuple[tuple[str, list[str], bool], ...] = (
15
+    ("qwen3-1.7b", [], False),
16
+    ("qwen3-1.7b-thinking", [], False),
17
+    ("qwen3-4b", [], False),
18
+    ("qwen3-8b", [], False),
19
+    ("llama-3.3-8b-instruct", ["--i-accept-license"], False),
20
+    ("phi-4-mini-reasoning", [], False),
21
+    ("gemma-2-2b-it", ["--i-accept-license"], False),
22
+    ("gemma-2-9b-it", ["--i-accept-license"], False),
23
+    ("mistral-small-3.1-24b-instruct", ["--multimodal"], True),
24
+    ("smollm3-3b", [], False),
25
+    ("olmo-2-7b-instruct", [], False),
26
+    ("mixtral-8x7b-instruct", [], False),
27
+    ("internvl3-2b", ["--multimodal"], True),
3728
 )
38
-def test_init_scaffolds_for_landed_registry_refresh_entries(
29
+
30
+
31
+@pytest.mark.parametrize(("base_key", "extra_flags", "expect_image_section"), SPRINT40_INIT_CASES)
32
+def test_init_scaffolds_for_every_registry_refresh_entry(
3933
     tmp_path: Path,
4034
     base_key: str,
4135
     extra_flags: list[str],
tests/integration/gate/test_mixtral_gate_smoke.pyadded
@@ -0,0 +1,54 @@
1
+"""Sprint 40 smoke proof that the Mixtral row still flows through Sprint 34 gate paths."""
2
+
3
+from __future__ import annotations
4
+
5
+from pathlib import Path
6
+from types import SimpleNamespace
7
+
8
+from dlm.base_models import BASE_MODELS
9
+from dlm.inference.gate import GateHandle, load_gate_handle
10
+from dlm.modality import modality_for
11
+from dlm.train.gate import GateTrainingSample, train_gate
12
+
13
+
14
+def _store(tmp_path: Path) -> SimpleNamespace:
15
+    return SimpleNamespace(root=tmp_path)
16
+
17
+
18
+def test_mixtral_text_moe_row_still_uses_text_gate_pipeline(tmp_path: Path) -> None:
19
+    import torch
20
+
21
+    spec = BASE_MODELS["mixtral-8x7b-instruct"]
22
+    dispatch = modality_for(spec)
23
+    assert spec.modality == "text-moe"
24
+    assert dispatch.accepts_images is False
25
+    assert dispatch.accepts_audio is False
26
+
27
+    store = _store(tmp_path)
28
+    samples: list[GateTrainingSample] = []
29
+    for _ in range(12):
30
+        samples.append(
31
+            GateTrainingSample(embedding=torch.ones(8) + 0.05 * torch.randn(8), adapter_name="a")
32
+        )
33
+        samples.append(
34
+            GateTrainingSample(embedding=-torch.ones(8) + 0.05 * torch.randn(8), adapter_name="b")
35
+        )
36
+
37
+    result = train_gate(
38
+        store,  # type: ignore[arg-type]
39
+        samples,
40
+        adapter_names=("a", "b"),
41
+        input_dim=8,
42
+        hidden_proj_dim=8,
43
+        steps=80,
44
+        lr=3e-3,
45
+        cold_start_floor=1,
46
+        batch_size=8,
47
+        seed=0,
48
+    )
49
+    assert result.mode == "trained"
50
+
51
+    handle = load_gate_handle(store)  # type: ignore[arg-type]
52
+    assert isinstance(handle, GateHandle)
53
+    assert handle.is_uniform is False
54
+    assert handle.metadata.adapter_names == ("a", "b")
tests/unit/base_models/test_audio_registry.pymodified
@@ -4,8 +4,8 @@ Mirrors `test_vl_registry.py` for the audio-language modality. Covers:
44
55
 - `qwen2-audio-7b-instruct` is present and has `modality="audio-language"`.
66
 - Its `AudioPreprocessorPlan` is pinned (16 kHz, 30 s, `<|AUDIO|>`, 750).
7
-- License is Apache-2.0 but the 7B weights are gated behind HF acceptance
8
-  and flagged non-redistributable (pack tarball size).
7
+- License is Apache-2.0 and the current HF row is no longer gated, so
8
+  the spec stays redistributable.
99
 - `modality="audio-language"` without a plan rejects at validate time;
1010
   text bases cannot carry an audio plan; VL bases cannot carry an audio
1111
   plan; audio bases cannot carry a VL plan.
@@ -48,10 +48,10 @@ class TestQwen2AudioRegistryEntry:
4848
         spec = BASE_MODELS["qwen2-audio-7b-instruct"]
4949
         assert spec.vl_preprocessor_plan is None
5050
51
-    def test_license_gated_not_redistributable(self) -> None:
51
+    def test_license_open_and_redistributable(self) -> None:
5252
         spec = BASE_MODELS["qwen2-audio-7b-instruct"]
53
-        assert spec.requires_acceptance is True
54
-        assert spec.redistributable is False
53
+        assert spec.requires_acceptance is False
54
+        assert spec.redistributable is True
5555
5656
     def test_architecture_is_audio_conditional_generation(self) -> None:
5757
         spec = BASE_MODELS["qwen2-audio-7b-instruct"]
tests/unit/base_models/test_registry.pymodified
@@ -73,11 +73,13 @@ class TestLicenseFields:
7373
             "qwen2.5-1.5b",
7474
             "qwen2.5-coder-1.5b",
7575
             "qwen3-1.7b",
76
+            "qwen3-1.7b-thinking",
7677
             "qwen3-4b",
7778
             "qwen3-8b",
7879
             "mixtral-8x7b-instruct",
7980
             "smollm3-3b",
8081
             "olmo-2-7b-instruct",
82
+            "qwen2-audio-7b-instruct",
8183
             "smollm2-135m",
8284
             "smollm2-360m",
8385
             "smollm2-1.7b",
@@ -137,7 +139,10 @@ class TestArchitectureShapes:
137139
             BASE_MODELS[k].size_gb_fp16 for k in ("qwen2.5-0.5b", "qwen2.5-1.5b", "qwen2.5-3b")
138140
         ]
139141
         assert qwen_sizes == sorted(qwen_sizes)
140
-        qwen3_sizes = [BASE_MODELS[k].size_gb_fp16 for k in ("qwen3-1.7b", "qwen3-4b", "qwen3-8b")]
142
+        qwen3_sizes = [
143
+            BASE_MODELS[k].size_gb_fp16
144
+            for k in ("qwen3-1.7b", "qwen3-1.7b-thinking", "qwen3-4b", "qwen3-8b")
145
+        ]
141146
         assert qwen3_sizes == sorted(qwen3_sizes)
142147
         smol_sizes = [
143148
             BASE_MODELS[k].size_gb_fp16 for k in ("smollm2-135m", "smollm2-360m", "smollm2-1.7b")
tests/unit/base_models/test_registry_2026.pymodified
@@ -37,6 +37,27 @@ class TestQwen3RegistryEntries:
3737
         assert spec.size_gb_fp16 == pytest.approx(16.0)
3838
3939
40
+class TestQwen3ThinkingRegistryEntry:
41
+    def test_entry_present(self) -> None:
42
+        assert "qwen3-1.7b-thinking" in BASE_MODELS
43
+
44
+    def test_reuses_live_qwen3_weights_with_reasoning_profile(self) -> None:
45
+        spec = BASE_MODELS["qwen3-1.7b-thinking"]
46
+        assert spec.hf_id == "Qwen/Qwen3-1.7B"
47
+        assert spec.architecture == "Qwen3ForCausalLM"
48
+        assert spec.template == "qwen3thinking"
49
+        assert spec.gguf_arch == "qwen3"
50
+        assert spec.tokenizer_pre == "qwen2"
51
+
52
+    def test_reasoning_profile_keeps_open_license_and_cooler_default(self) -> None:
53
+        spec = BASE_MODELS["qwen3-1.7b-thinking"]
54
+        assert spec.license_spdx == "Apache-2.0"
55
+        assert spec.requires_acceptance is False
56
+        assert spec.redistributable is True
57
+        assert spec.reasoning_tuned is True
58
+        assert spec.suggested_prompt_temperature == pytest.approx(0.6)
59
+
60
+
4061
 class TestLlama33RegistryEntry:
4162
     def test_entry_present(self) -> None:
4263
         assert "llama-3.3-8b-instruct" in BASE_MODELS
@@ -212,28 +233,23 @@ class TestMixtralRegistryEntry:
212233
         assert spec.recommended_seq_len == 2048
213234
214235
215
-class TestStaleSprintDraftRows:
216
-    def test_qwen3_thinking_is_not_a_separate_registry_row(self) -> None:
217
-        """Upstream Qwen3-1.7B ships hybrid thinking in one model.
218
-
219
-        Sprint 40's draft listed a separate `qwen3-1.7b-thinking`
220
-        entry, but the live upstream contract exposes thinking mode as
221
-        a switch on `Qwen/Qwen3-1.7B` itself. Keep the registry honest:
222
-        reasoning defaults belong on the real base row, not a fake key.
223
-        """
224
-        assert "qwen3-1.7b-thinking" not in BASE_MODELS
225
-
226
-    def test_internvl3_not_shipped_until_remote_code_contract_is_pinned(self) -> None:
227
-        """Guard against copying the stale sprint draft into the registry.
228
-
229
-        The live `OpenGVLab/InternVL3-2B` model card still documents
230
-        `trust_remote_code=True`, and on the current stack the whole
231
-        InternVL family still exposes a tokenizer-only `AutoProcessor`
232
-        rather than a complete image processor. Upstream also expands
233
-        `<image>` into repeated `<IMG_CONTEXT>` spans and threads
234
-        `image_flags` through the forward pass. Adding InternVL3 later
235
-        is fine, but it needs an honest runtime contract instead of
236
-        assuming the old "cleaner than InternVL2" sprint note is still
237
-        true.
238
-        """
239
-        assert "internvl3-2b" not in BASE_MODELS
236
+class TestInternVL3RegistryEntry:
237
+    def test_entry_present(self) -> None:
238
+        assert "internvl3-2b" in BASE_MODELS
239
+
240
+    def test_entry_keeps_remote_code_contract_explicit(self) -> None:
241
+        spec = BASE_MODELS["internvl3-2b"]
242
+        assert spec.hf_id == "OpenGVLab/InternVL3-2B"
243
+        assert spec.architecture == "InternVLChatModel"
244
+        assert spec.template == "internvl2"
245
+        assert spec.trust_remote_code is True
246
+
247
+    def test_entry_is_registry_visible_but_not_pretending_runtime_is_generic(self) -> None:
248
+        spec = BASE_MODELS["internvl3-2b"]
249
+        assert spec.license_spdx == "Apache-2.0"
250
+        assert spec.requires_acceptance is False
251
+        assert spec.redistributable is True
252
+        assert spec.modality == "vision-language"
253
+        assert spec.vl_preprocessor_plan is not None
254
+        assert spec.vl_preprocessor_plan.resize_policy == "dynamic"
255
+        assert spec.vl_preprocessor_plan.image_token == "<image>"
tests/unit/base_models/test_schema.pymodified
@@ -89,7 +89,17 @@ class TestTargetModules:
8989
 class TestLiteralConstraints:
9090
     @pytest.mark.parametrize(
9191
         "template",
92
-        ["chatml", "gemma2", "smollm3", "olmo2", "llama3", "phi3", "phi4mini", "mistral"],
92
+        [
93
+            "chatml",
94
+            "qwen3thinking",
95
+            "gemma2",
96
+            "smollm3",
97
+            "olmo2",
98
+            "llama3",
99
+            "phi3",
100
+            "phi4mini",
101
+            "mistral",
102
+        ],
93103
     )
94104
     def test_template_literals_accepted(self, template: str) -> None:
95105
         spec = _minimal(template=template)
tests/unit/base_models/test_vl_registry.pymodified
@@ -143,6 +143,7 @@ _VL_BASE_KEYS: tuple[str, ...] = (
143143
     "paligemma-3b-mix-224",
144144
     "qwen2-vl-2b-instruct",
145145
     "internvl2-2b",
146
+    "internvl3-2b",
146147
     "mistral-small-3.1-24b-instruct",
147148
 )
148149
@@ -160,7 +161,7 @@ class TestAllVlBasesShipModalityInvariants:
160161
         assert spec.vl_preprocessor_plan is not None
161162
         # Pinned identity fields — each one is part of the cache key,
162163
         # so a silent default would silently invalidate caches.
163
-        assert spec.vl_preprocessor_plan.resize_policy == "fixed"
164
+        assert spec.vl_preprocessor_plan.resize_policy in {"fixed", "dynamic"}
164165
         assert spec.vl_preprocessor_plan.num_image_tokens > 0
165166
166167
     @pytest.mark.parametrize("key", _VL_BASE_KEYS)
@@ -241,28 +242,56 @@ class TestInternVL2RegistryEntry:
241242
         assert BASE_MODELS["internvl2-2b"].template == "internvl2"
242243
243244
245
+class TestInternVL3RegistryEntry:
246
+    """Sprint 40 refresh: InternVL3 lands with an explicit runtime caveat."""
247
+
248
+    def test_entry_present(self) -> None:
249
+        assert "internvl3-2b" in BASE_MODELS
250
+
251
+    def test_apache_permissive(self) -> None:
252
+        spec = BASE_MODELS["internvl3-2b"]
253
+        assert spec.license_spdx == "Apache-2.0"
254
+        assert spec.requires_acceptance is False
255
+        assert spec.redistributable is True
256
+
257
+    def test_dynamic_preprocessing_plan_is_pinned(self) -> None:
258
+        spec = BASE_MODELS["internvl3-2b"]
259
+        plan = spec.vl_preprocessor_plan
260
+        assert plan is not None
261
+        assert plan.target_size == (448, 448)
262
+        assert plan.resize_policy == "dynamic"
263
+        assert plan.image_token == "<image>"
264
+        assert plan.num_image_tokens == 256
265
+
266
+    def test_architecture_and_template(self) -> None:
267
+        spec = BASE_MODELS["internvl3-2b"]
268
+        assert spec.architecture == "InternVLChatModel"
269
+        assert spec.template == "internvl2"
270
+        assert spec.trust_remote_code is True
271
+
272
+
244273
 class TestDistinctVlBases:
245274
     """The VL bases occupy distinct rows with no silent duplicates."""
246275
247276
     def test_all_keys_unique(self) -> None:
248
-        assert len(set(_VL_BASE_KEYS)) == 4
277
+        assert len(set(_VL_BASE_KEYS)) == 5
249278
250279
     def test_hf_ids_distinct(self) -> None:
251280
         hf_ids = {BASE_MODELS[k].hf_id for k in _VL_BASE_KEYS}
252
-        assert len(hf_ids) == 4
281
+        assert len(hf_ids) == 5
253282
254283
     def test_image_tokens_distinct_per_base(self) -> None:
255
-        """Each VL base uses its native image-token string.
284
+        """VL rows pin their native placeholder tokens explicitly.
256285
257
-        Silently sharing a placeholder across bases would break the
258
-        cache-key invariant in vl_cache.py (cache key includes the
259
-        token via processor_sha256).
286
+        Some families legitimately reuse the same surface token
287
+        (`<image>`), so this checks the concrete set rather than
288
+        forcing uniqueness for uniqueness' sake.
260289
         """
261290
         tokens = {
262291
             BASE_MODELS[k].vl_preprocessor_plan.image_token  # type: ignore[union-attr]
263292
             for k in _VL_BASE_KEYS
264293
         }
265
-        assert len(tokens) == 4
294
+        assert tokens == {"<image>", "<|image_pad|>", "<IMG_CONTEXT>", "[IMG]"}
266295
267296
268297
 class TestCountVlRegistryEntries:
@@ -270,7 +299,7 @@ class TestCountVlRegistryEntries:
270299
271300
     def test_at_least_four_vl_bases_registered(self) -> None:
272301
         vl_count = sum(1 for s in BASE_MODELS.values() if s.modality == "vision-language")
273
-        assert vl_count >= 4
302
+        assert vl_count >= 5
274303
275304
276305
 class TestTrustRemoteCodeOptIn:
@@ -292,6 +321,9 @@ class TestTrustRemoteCodeOptIn:
292321
         is defined in the model repo, not in transformers."""
293322
         assert BASE_MODELS["internvl2-2b"].trust_remote_code is True
294323
324
+    def test_internvl3_opts_in(self) -> None:
325
+        assert BASE_MODELS["internvl3-2b"].trust_remote_code is True
326
+
295327
     def test_text_bases_default_false(self) -> None:
296328
         """None of the text bases opt into trust_remote_code."""
297329
         for key, spec in BASE_MODELS.items():
tests/unit/doc/test_v12_migrator.pyadded
@@ -0,0 +1,36 @@
1
+"""Named Sprint 40 closeout checks for the v12 → v13 migrator."""
2
+
3
+from __future__ import annotations
4
+
5
+from typing import Any
6
+
7
+from dlm.doc.migrations.v12 import migrate
8
+from dlm.doc.schema import DlmFrontmatter
9
+
10
+_VALID_ULID = "01HZ4X7TGZM3J1A2B3C4D5E6F7"
11
+
12
+
13
+def test_v12_migrator_is_identity_for_existing_frontmatter() -> None:
14
+    raw: dict[str, Any] = {
15
+        "dlm_id": _VALID_ULID,
16
+        "base_model": "smollm2-135m",
17
+        "dlm_version": 12,
18
+        "training": {"audio": {"auto_resample": True}, "lora_r": 16},
19
+    }
20
+    out = migrate(raw)
21
+    assert out == raw
22
+    assert out is not raw
23
+
24
+
25
+def test_v12_migrator_output_validates_as_v13() -> None:
26
+    raw: dict[str, Any] = {
27
+        "dlm_id": _VALID_ULID,
28
+        "base_model": "smollm2-135m",
29
+        "dlm_version": 12,
30
+        "training": {"audio": {"auto_resample": True}},
31
+    }
32
+    out = migrate(raw)
33
+    out["dlm_version"] = 13
34
+    fm = DlmFrontmatter.model_validate(out)
35
+    assert fm.dlm_version == 13
36
+    assert fm.training.audio.auto_resample is True
tests/unit/export/ollama/test_template_registry.pymodified
@@ -13,9 +13,10 @@ from dlm.export.ollama.template_registry import (
1313
1414
1515
 class TestRegistryCoverage:
16
-    def test_all_eight_dialects_registered(self) -> None:
16
+    def test_all_nine_dialects_registered(self) -> None:
1717
         assert set(registered_dialects()) == {
1818
             "chatml",
19
+            "qwen3thinking",
1920
             "gemma2",
2021
             "smollm3",
2122
             "olmo2",
@@ -27,7 +28,17 @@ class TestRegistryCoverage:
2728
2829
     @pytest.mark.parametrize(
2930
         "dialect",
30
-        ["chatml", "gemma2", "smollm3", "olmo2", "llama3", "phi3", "phi4mini", "mistral"],
31
+        [
32
+            "chatml",
33
+            "qwen3thinking",
34
+            "gemma2",
35
+            "smollm3",
36
+            "olmo2",
37
+            "llama3",
38
+            "phi3",
39
+            "phi4mini",
40
+            "mistral",
41
+        ],
3142
     )
3243
     def test_each_template_file_exists(self, dialect: str) -> None:
3344
         row = get_template(dialect)
@@ -37,7 +48,17 @@ class TestRegistryCoverage:
3748
3849
     @pytest.mark.parametrize(
3950
         "dialect",
40
-        ["chatml", "gemma2", "smollm3", "olmo2", "llama3", "phi3", "phi4mini", "mistral"],
51
+        [
52
+            "chatml",
53
+            "qwen3thinking",
54
+            "gemma2",
55
+            "smollm3",
56
+            "olmo2",
57
+            "llama3",
58
+            "phi3",
59
+            "phi4mini",
60
+            "mistral",
61
+        ],
4162
     )
4263
     def test_each_has_default_stops(self, dialect: str) -> None:
4364
         row = get_template(dialect)
@@ -47,6 +68,7 @@ class TestRegistryCoverage:
4768
         ("dialect", "required"),
4869
         [
4970
             ("chatml", {"<|im_end|>", "<|im_start|>"}),
71
+            ("qwen3thinking", {"<|im_end|>", "<|im_start|>"}),
5072
             ("gemma2", {"<end_of_turn>", "<start_of_turn>"}),
5173
             ("smollm3", {"<|im_end|>", "<|im_start|>"}),
5274
             ("olmo2", {"<|endoftext|>", "<|user|>", "<|assistant|>"}),
@@ -89,6 +111,14 @@ class TestDialectShapes:
89111
         assert "<end_of_turn>" in text
90112
         assert "model" in text
91113
114
+    def test_qwen3thinking_keeps_chatml_markers_with_reasoning_defaults(self) -> None:
115
+        text = load_template_text("qwen3thinking")
116
+        assert "<|im_start|>" in text
117
+        assert "<|im_end|>" in text
118
+        row = get_template("qwen3thinking")
119
+        assert row.default_temperature == pytest.approx(0.6)
120
+        assert row.default_top_p == pytest.approx(0.95)
121
+
92122
     def test_smollm3_has_reasoning_system_prompt(self) -> None:
93123
         text = load_template_text("smollm3")
94124
         assert "<|im_start|>system" in text
tests/unit/export/test_mixtral_template.pyadded
@@ -0,0 +1,38 @@
1
+"""Sprint 40 closeout checks for the Mixtral template row."""
2
+
3
+from __future__ import annotations
4
+
5
+import json
6
+from pathlib import Path
7
+
8
+from dlm.base_models import BASE_MODELS
9
+from dlm.export.ollama.modelfile import ModelfileContext, render_modelfile
10
+from dlm.export.plan import ExportPlan
11
+
12
+
13
+def _adapter_dir(tmp_path: Path) -> Path:
14
+    adapter = tmp_path / "adapter"
15
+    adapter.mkdir()
16
+    (adapter / "tokenizer_config.json").write_text(
17
+        json.dumps({"eos_token": "</s>", "added_tokens_decoder": {}}),
18
+        encoding="utf-8",
19
+    )
20
+    return adapter
21
+
22
+
23
+def test_mixtral_registry_row_renders_through_mistral_template(tmp_path: Path) -> None:
24
+    spec = BASE_MODELS["mixtral-8x7b-instruct"]
25
+    text = render_modelfile(
26
+        ModelfileContext(
27
+            spec=spec,
28
+            plan=ExportPlan(quant="Q4_K_M", merged=False),
29
+            adapter_dir=_adapter_dir(tmp_path),
30
+            base_gguf_name="base.gguf",
31
+            adapter_gguf_name="adapter.gguf",
32
+            dlm_id="01TEST",
33
+            adapter_version=1,
34
+        )
35
+    )
36
+    assert spec.modality == "text-moe"
37
+    assert "[INST]" in text
38
+    assert 'PARAMETER stop "[INST]"' in text
tests/unit/export/test_phi4_template.pyadded
@@ -0,0 +1,37 @@
1
+"""Sprint 40 closeout checks for the Phi-4 reasoning template row."""
2
+
3
+from __future__ import annotations
4
+
5
+import json
6
+from pathlib import Path
7
+
8
+from dlm.base_models import BASE_MODELS
9
+from dlm.export.ollama.modelfile import ModelfileContext, render_modelfile
10
+from dlm.export.plan import ExportPlan
11
+
12
+
13
+def _adapter_dir(tmp_path: Path) -> Path:
14
+    adapter = tmp_path / "adapter"
15
+    adapter.mkdir()
16
+    (adapter / "tokenizer_config.json").write_text(
17
+        json.dumps({"eos_token": "<|end|>", "added_tokens_decoder": {}}),
18
+        encoding="utf-8",
19
+    )
20
+    return adapter
21
+
22
+
23
+def test_phi4_reasoning_template_keeps_phi_system_preamble_and_stops(tmp_path: Path) -> None:
24
+    text = render_modelfile(
25
+        ModelfileContext(
26
+            spec=BASE_MODELS["phi-4-mini-reasoning"],
27
+            plan=ExportPlan(quant="Q4_K_M", merged=False),
28
+            adapter_dir=_adapter_dir(tmp_path),
29
+            base_gguf_name="base.gguf",
30
+            adapter_gguf_name="adapter.gguf",
31
+            dlm_id="01TEST",
32
+            adapter_version=1,
33
+        )
34
+    )
35
+    assert "Your name is Phi, an AI math expert developed by Microsoft." in text
36
+    assert 'PARAMETER stop "<|assistant|>"' in text
37
+    assert "PARAMETER temperature 0.6" in text
tests/unit/export/test_qwen3_template.pyadded
@@ -0,0 +1,37 @@
1
+"""Sprint 40 closeout checks for the Qwen3 reasoning-template row."""
2
+
3
+from __future__ import annotations
4
+
5
+import json
6
+from pathlib import Path
7
+
8
+from dlm.base_models import BASE_MODELS
9
+from dlm.export.ollama.modelfile import ModelfileContext, render_modelfile
10
+from dlm.export.plan import ExportPlan
11
+
12
+
13
+def _adapter_dir(tmp_path: Path) -> Path:
14
+    adapter = tmp_path / "adapter"
15
+    adapter.mkdir()
16
+    (adapter / "tokenizer_config.json").write_text(
17
+        json.dumps({"eos_token": "<|im_end|>", "added_tokens_decoder": {}}),
18
+        encoding="utf-8",
19
+    )
20
+    return adapter
21
+
22
+
23
+def test_qwen3_thinking_row_uses_distinct_reasoning_template_defaults(tmp_path: Path) -> None:
24
+    text = render_modelfile(
25
+        ModelfileContext(
26
+            spec=BASE_MODELS["qwen3-1.7b-thinking"],
27
+            plan=ExportPlan(quant="Q4_K_M", merged=False),
28
+            adapter_dir=_adapter_dir(tmp_path),
29
+            base_gguf_name="base.gguf",
30
+            adapter_gguf_name="adapter.gguf",
31
+            dlm_id="01TEST",
32
+            adapter_version=1,
33
+        )
34
+    )
35
+    assert "PARAMETER temperature 0.6" in text
36
+    assert "PARAMETER top_p 0.95" in text
37
+    assert "<|im_start|>assistant" in text