Choose a task below. Each recipe documents its working directory, model/data
requirements, commands, and expected output. Model weights and datasets are
separate downloads; check their licenses before use.
See the README recipe audit for what was checked,
fixed, and actually executed on the current checkout.
Generation and serving
| Workflow | Recipe |
|---|
| Direct inference from a supported Hugging Face model ID | HF inference |
| Cloud Qwen Chat with automatic preparation and authentication | RunPod deployment |
| FastAPI server, native inference, OpenAI-style chat, Anthropic-style Messages | Server |
| Qwen3.8-27B text inference, MTP verification, and baseline comparison | Qwen3.8 / MTP |
| Claude Code CLI backed by local Qwen3.8 | Claude Code |
| Qwen3.5 text, vision, streaming, and serving | Qwen3.5 |
| Llama inference and fine-tuning | Llama2, Llama3 |
| Mistral and Mixtral inference | Mistral, Mixtral |
Translation and training
Multimodal models
Contributing a recipe
Use a single README.md so the documentation build can include the recipe.
Specify the working directory, exact checkpoint, dependencies, expected output,
and limitations. Keep scripts configurable, avoid machine-specific paths, and
state what was actually tested. Link the recipe here and from the main README
when it is a useful starting point. Respect model and dataset licenses.