Ornith 1.0: 9B vs 35B local coding model tests

This first look compares Ornith 1.0's 9B dense and 35B mixture-of-experts models as local LLM options for coding. The speaker describes both as Qwen fine-tunes. At filming, the listed Gemma 4-based 31B model was unavailable on Hugging Face, and the speaker lacked hardware to run the 397B version.

The 9B runs at Q8 quantization in LM Studio on a laptop 5090 with 24 GB of video memory. Most 35B tests use OpenCode with an unquantized model served through vLLM on a 6000 Pro system. These different configurations limit the comparison; the hardware shown is a test setup, not a stated minimum requirement.

Tasks include a browser desktop, subway scenes and FPS games, a watch storefront with 3D assets, and image-based code generation. The speaker favors the 35B's desktop and watch website, though its file manager is static and its repaired subway game still has lag and broken enemy-damage logic. The C++ racing task is stopped without a usable result. A separate Q8 35B speedboat test also disappoints.

Moving the 9B into OpenCode helps repair its image-to-3D output, which reproduces parts of an arcade machine. Its subway game and watch assets remain broken. The speaker prefers the 35B in this limited test but does not establish an improvement over the base Qwen models. Possible training exposure to familiar prompts is another caveat.