gpt-oss-120b / 20b

OpenAI · Released: unverified (Ollama lists cloud tags as ~1 year old) · Last verified 2026-10-16

Apache-2.0 MoE models shipped natively in MXFP4. Uses the Harmony format; many early Jinja templates rendered it wrong.

Recommended quants

Native MXFP4 (120b ~65 GB, 20b ~14 GB in Ollama)
ollama gpt-oss:120b / :20b
Ollama supports MXFP4 natively.
Unsloth GGUF (F16 / UD-Q2_K_XL etc.)
unsloth/gpt-oss-120b-GGUF, unsloth/gpt-oss-20b-GGUF
Unsloth: sub-F16 quants change little because most weights stay full precision.

Chat template Caveats

Unsloth found several differences between common Jinja templates and OpenAI's Harmony renderer (tool-call escaping, extra newlines, wrong channel for tool thoughts, missing final channel).

Fix: Use Unsloth uploads (fixes partly upstreamed per Unsloth) or render with Harmony. Run llama.cpp with --jinja.

Sampling settings

All tasks (OpenAI)	`temp 1.0, top_p 1.0, top_k 0 (or try 100)`
Context	`min 16,384 recommended; max 131,072`
Reasoning effort	`low / medium / high`

Runtimes

vLLM	Unverified	Not checked in an opened source.
llama.cpp	Works	Documented by Unsloth; -ot ".ffn_.*_exps.=CPU" to offload MoE layers.
Ollama	Works	Official gpt-oss tags, native MXFP4.
LM Studio	Works	Listed as a supported way to run 20b by Unsloth.
MLX	Unverified

Tool calling in agent tools

Claude Code (local endpoint)	Unverified	Ollama documents `ollama launch claude --model gpt-oss`.
OpenCode	Unverified	Ollama documents `ollama launch opencode --model gpt-oss`.
Cline	Unverified
Aider	Unverified

Known issues

Unsloth: Jinja vs Harmony template differences and precision issues on float16 GPUs

Sources checked

Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)