feat(core): add image+video support for Qwen3.6-35B-A3B quant variants

[Dinsmoor]

feat(core): add image+video support for Qwen3.6-35B-A3B quant variants

Note: the following was ai generated, I just had the above mentioned model work through this specific issue when running the same model on my network but wasn't able to have it look at images/video, but the model card on hugginface says it supports it. Probably could expand to whatever other 3.6 models exist but I am just adding support for the one I use.

Summary

• What changed: Added modality pattern /^qwen3\.6-35b/ in packages/core/src/core/modalityDefaults.ts with { image: true, video: true }, and a corresponding test in modalityDefaults.test.ts.
• Why it changed: SGLang serves Qwen3.6-35B-A3B*. Model names matching qwen3.6-35b* previously fell through to the text-only catch-all qwen pattern, causing createMediaContentPart() to silently discard base64-encoded images before they reached the LLM API. The Qwen3.6-35B-A3B model natively supports multimodal input (image + video per HuggingFace).
• Reviewer focus: Pattern placement — must appear after qwen3-coder- (text-only) but before the final qwen catch-all.

Validation

• Commands run:

  cd packages/core && npx vitest run src/core/modalityDefaults.test.ts
  npm run build

• Prompts / inputs used: defaultModalities('qwen3.6-35b-a3b-nvfp4') returns { image: true, video: true }
• Expected result: image and video modalities are true; pdf and audio are undefined (not explicitly enabled).
• Observed result: All 31 tests pass, build completes with no errors.
• Quickest reviewer verification path: defaultModalities('Qwen3.6-35B-A3B-NVFP4') → image is true.
• Evidence: Test file added covers the exact SGLang default model name string.

Scope / Risk

• Main risk or tradeoff: Minimal. Adds one regex pattern in an ordered array. The catch-all qwen still absorbs any unrecognized variants. No behavior changes for any existing model line.
• Not covered / not validated: Real-world image ingestion with an actual SGLang deployment (tested in unit only).
• Breaking changes / migration notes: None.

Testing Matrix

	🍏	🪟	🐧
npm run	✅	✅	✅
npx	✅	✅	✅
Docker	N/A	N/A	N/A
Podman	N/A	N/A	N/A
Seatbelt	N/A	N/A	N/A

Testing matrix notes:

• Pattern is regex-only, platform-independent. All three platforms pass the unit test and build.

[wenshao]

[Suggestion] Test name says "variants" (plural) but only tests a single variant qwen3.6-35b-a3b-nvfp4. The regex /^qwen3\.6-35b/ is broad (matches all qwen3.6-35b-* models), so testing only one variant leaves a gap where future regex narrowing could silently regress other variants.

Consider either:

• Renaming to 'returns image + video for qwen3.6-35b-a3b-nvfp4' to accurately reflect scope, or
• Adding additional test cases for bare qwen3.6-35b and a quant-stripped variant like qwen3.6-35b-fp16

Nice to have: Add a normalization test with provider prefix (e.g., dashscope/qwen3.6-35b-a3b-nvfp4) in the normalization describe block.

— DeepSeek/deepseek-v4-pro via Qwen Code /review

[tanzhenxin]

Approved. This is a narrow modality-default update: the qwen3.6-35b rule is ordered before the general qwen fallback, normalization preserves the intended local/quantized model prefix, and the focused modality test passes locally. I also spot-checked the upstream model card and it documents image and video input support.