Run new open models right, the first time.
One card per model: which quant, which chat template fix, which sampling settings, and what actually works in vLLM, llama.cpp, Ollama, LM Studio, MLX and your coding agent. Every claim links to its source.
Latest cards
Gemma 4 (26B-A4B / 31B)
Google DeepMind · verified 2026-10-16
Apache-2.0 multimodal family. 26B-A4B MoE for speed, 31B dense for quality. New thinking control tokens and a strict multi-turn rule.
Template Works · 4 works · 0 caveats · 5 unverified
gpt-oss-120b / 20b
OpenAI · verified 2026-10-16
Apache-2.0 MoE models shipped natively in MXFP4. Uses the Harmony format; many early Jinja templates rendered it wrong.
Template Caveats · 3 works · 0 caveats · 6 unverified
Qwen3.6-27B
Alibaba Qwen · verified 2026-10-16
Dense 27B hybrid-thinking model, popular for local agentic coding. Most pain is chat-template / tool-call handling in agent loops.
Template Caveats · 3 works · 3 caveats · 3 unverified
Qwen3.6-35B-A3B
Alibaba Qwen · verified 2026-10-16
MoE sibling of Qwen3.6-27B (3B active). Faster decode; same template family and same tool-call caveats.
Template Caveats · 3 works · 1 caveats · 5 unverified
Qwen3.8-27B
Alibaba Qwen · verified 2026-10-16
Newest dense Qwen with reasoning_effort control. Biggest gotcha: official template defaults to xhigh reasoning, which can burn the token budget.
Template Caveats · 2 works · 0 caveats · 7 unverified