feat: add stronger local model

[AlexanderMelde]

What does this PR do?

This pull request adds support for several new Qwen model variants to the model loader, including their configuration, context sizes, and metadata. These additions expand the available models, especially for users needing efficient models with lower VRAM requirements or improved reasoning capabilities.

New model support and configuration:

• Added three new Qwen models to the F16_TO_F32_MODEL_MAP for automatic mapping between quantization formats: Qwen2.5-3B-Instruct-q4f16_1-MLC, Qwen3.5-4B-Instruct-q4f16_1-MLC, and Qwen3.5-35B-A3B-q4f16_1-MLC.
• Registered new context sizes (32k tokens) for both f16 and f32 versions of the new Qwen models in MODEL_CONTEXT_SIZES.

Model metadata and capabilities:

• Added detailed metadata for three new models in the ModelLoader class:
  • Qwen3.5-35B-A3B-q4f16_1-MLC: Large MoE model with high reasoning quality and low VRAM usage, marked as recommended.
  • Qwen3.5-4B-Instruct-q4f16_1-MLC: Balanced model for agentic tasks.
  • Qwen2.5-3B-Instruct-q4f16_1-MLC: Reliable small model for structured tasks.

Type

• New ability
• New workflow
• Bug fix
• Enhancement
• Docs

How to test

Screenshots (if UI changes)

[pluginslab]

Closing for now. Qwen 3.5 uses a hybrid Gated DeltaNet architecture that TVM/MLC cannot compile yet, so these model IDs likely do not exist in the current @mlc-ai/web-llm prebuilt list. Selecting them from the dropdown would fail at runtime.

We will revisit this when Qwen 3.5 becomes available in WebLLM. Thanks for the contribution!