gpt-oss-20b ยท GGUF (MXFP4)
GGUF build of openai/gpt-oss-20b for llama.cpp / Ollama / LM Studio. Converted by axteam with llama.cpp. The chat template (OpenAI harmony format) is embedded in the GGUF.
| File | Quant | Size | Notes |
|---|---|---|---|
gpt-oss-20b-GGUF-MXFP4.gguf |
MXFP4 | 13.8 GB | native 4-bit MoE weights (as released by OpenAI) |
gpt-oss-20b is natively MXFP4 โ the expert weights ship in 4-bit, so this single file is the faithful, full-quality build (no separate Q4/Q8 needed). OpenAI states gpt-oss-20b runs within 16 GB of memory; it was run here on an NVIDIA B200 (2026-11-02, see Requirements). Architecture: gpt-oss ยท Mixture-of-Experts (~21B total, ~3.6B active). 128k context, adjustable reasoning effort.
Quantization method
No quantization was done in this repo. OpenAI's original card states that the models were post-trained with MXFP4 quantization of the MoE weights, so the 4-bit expert weights come from OpenAI's own training, not from a separate post-training quantization (PTQ) step. llama.cpp copies those MXFP4 expert weights into GGUF as-is, with no re-quantization and no retraining by axteam. The attention, token-embedding and output weights stay BF16 as in the original; norms, biases and the router are stored as F32. The precision matches OpenAI's original release. No quality figure is published for this file: there is no other build to compare it with, and its plain-text perplexity (measured on 2026-11-02) is not a meaningful quality measure for this harmony chat-format model.
Requirements
| llama.cpp | build โฅ b6325 โ architecture gpt-oss needs โฅ b6096, and the commands on this card use -fa on (โฅ b6325). Run here on 2026-11-02 with b12005 (commit e0f142d, CUDA build) on an NVIDIA B200: the file loaded with all layers on the GPU and answered two short chat questions (capital of France / of Korea) correctly (โParisโ, โ์์ธโ). |
| Memory | the chosen GGUF file (sizes in the table) plus KV cache for the context (-c); --cache-type-k/v q8_0 -fa on halves the KV cache. |
| Input | text. |
| Ollama / LM Studio | not tested here. |
Run with llama.cpp
Use a recent llama.cpp build (gpt-oss support, Aug-2025 or later).
# chat server (harmony template is embedded; --jinja uses it)
llama-server -m gpt-oss-20b-GGUF-MXFP4.gguf -c 8192 --jinja
# single answer and exit (-st)
llama-cli -m gpt-oss-20b-GGUF-MXFP4.gguf --jinja -p "Explain MoE in one sentence." -st
Reasoning effort can be set in the system prompt (Reasoning: low|medium|high).
Reduce KV-cache memory (optional, run-time)
KV-cache type is a run-time choice in llama.cpp (not baked into the file):
llama-server -m gpt-oss-20b-GGUF-MXFP4.gguf -c 8192 --jinja \
--cache-type-k q8_0 --cache-type-v q8_0 -fa on
Ollama
ollama run hf.co/axteam/gpt-oss-20b-GGUF
License
Apache-2.0, following the original openai/gpt-oss-20b. Weights, tokenizer, and chat template belong to OpenAI. This repo only re-packages them as GGUF.
ํ๊ตญ์ด ์์ฝ
openai/gpt-oss-20b์ GGUF ๋ณํ ๋น๋ (llama.cppยทOllamaยทLM Studio์ฉ) ยท ์ ์ axteam
๊ฐ์
| ํญ๋ชฉ | ๋ด์ฉ |
|---|---|
| ์๋ณธ ๋ชจ๋ธ | openai/gpt-oss-20b |
| ํ์ | GGUF (llama.cpp) |
| ๋ณํ ๋๊ตฌ | llama.cpp convert_hf_to_gguf.py (OpenAI ์๋ณธ MXFP4 ๊ฐ์ค์น๋ฅผ ๊ทธ๋๋ก ์ฎ๊น, llama-quantize ๋ฏธ์ฌ์ฉ) |
| ๋ชจ๋ธ ๊ตฌ์กฐ | MoE (์ด ์ฝ 21B ์ค ์ฝ 3.6B ํ์ฑ) ยท ์ปจํ ์คํธ 128k |
| ํ์ผ ์ | 1๊ฐ |
์์ํ ์์ธ
- ์์ํ ์ ์ฐจ: ์ด ๋ฆฌํฌ์์ ์๋ก ์์ํํ์ง ์์ โ OpenAI ์๋ณธ ์นด๋ ๊ธฐ์ค MoE ์ ๋ฌธ๊ฐ ๊ฐ์ค์น๋ MXFP4๋ก ์์ํํ ์ํ๋ก ์ฌํ ํ์ต(post-training)๋จ, llama.cpp๊ฐ ๊ทธ MXFP4 ๊ฐ์ค์น๋ฅผ ์ฌ์์ํยท์ฌํ์ต ์์ด GGUF๋ก ์ฎ๊น (์ดํ ์ ยท์๋ฒ ๋ฉยท์ถ๋ ฅ ๊ฐ์ค์น๋ ์๋ณธ์ฒ๋ผ BF16, ์ ๊ทํ์ธตยทbiasยท๋ผ์ฐํฐ๋ F32๋ก ์ ์ฅ)
- chat template: GGUF ํ์ผ์ ๋ด์ฅ (OpenAI harmony ํ์,
--jinja๋ก ์ฌ์ฉ) - ํ์ง: ์์น ๋ฏธ๊ธฐ์ฌ โ ๋น๊ตํ ๋ค๋ฅธ ํ์ผ์ด ์๊ณ , 2026-11-02์ ์ธก์ ํ ์ผ๋ฐ ํ ์คํธ PPL์ harmony ์ฑํ ํ์ ๋ชจ๋ธ์ด๋ผ ํ์ง ์งํ๋ก ์๋ฏธ๊ฐ ์์ด ์ฃ์ง ์์ ยท ์ฌ์์ํ ์์ด OpenAI ๋ฐฐํฌ๋ณธ๊ณผ ๊ฐ์ ์ ๋ฐ๋
์ ๊ณต ํ์ผ
| ํ์ผ | ํฌ๊ธฐ | ์ค๋ช |
|---|---|---|
gpt-oss-20b-GGUF-MXFP4.gguf |
13.8 GB | OpenAI๊ฐ ๋ฐฐํฌํ ๊ทธ๋๋ก์ 4๋นํธ MoE ๊ฐ์ค์น |
์๊ตฌ์ฌํญ
- llama.cpp: b6325 ์ด์ (
gpt-oss๊ตฌ์กฐ ์ง์ b6096๋ถํฐ, ์์ ๋ช ๋ น์-fa on์ต์ b6325๋ถํฐ) - ์คํ ํ์ธ: 2026-11-02 b12005(์ปค๋ฐ e0f142d) CUDA ๋น๋๋ก NVIDIA B200์์ ์ ์ธต์ GPU์ ์ฌ๋ ค ์คํ โ ์งง์ ์ง๋ฌธ 2๊ฐ(ํ๋์คยท๋ํ๋ฏผ๊ตญ ์๋)์ ๋ฐ๋ฅด๊ฒ ๋ตํจ(โParisโ, โ์์ธโ)
- ๋ฉ๋ชจ๋ฆฌ: ํ์ผ 13.8 GB + ์ปจํ
์คํธ ๊ธธ์ด(
-c)์ ๋น๋กํ๋ KV ์บ์ โ OpenAI ์๋ด ๊ธฐ์ค 16 GB ๋ฉ๋ชจ๋ฆฌ์์ ์คํ ๊ฐ๋ฅ (์ฌ๊ธฐ์๋ B200์์๋ง ์คํ โ 16 GB ๋ฉ๋ชจ๋ฆฌ ์คํ์ ๋ฏธํ์ธ, KV ์บ์๋ฅผ ์ค์ด๋ ๋ฐฉ๋ฒ์ ์๋ ์ฌ์ฉ๋ฒ ์ฐธ๊ณ ) - ์ ๋ ฅ: ํ ์คํธ
- OllamaยทLM Studio: ์ ์ ์ธก์์ ์คํ ๋ฏธํ์ธ
์ฌ์ฉ๋ฒ
- ํ์ผ ๋ฐ๊ธฐ:
hf download axteam/gpt-oss-20b-GGUF gpt-oss-20b-GGUF-MXFP4.gguf --local-dir .(์๋-m๋ช ๋ น์ ํ์) - ์๋ฒ ์คํ:
llama-server -m gpt-oss-20b-GGUF-MXFP4.gguf -c 8192 --jinja - ํ ๋ฒ ์ง๋ฌธ (๋ต๋ณ ํ ์ข
๋ฃ):
llama-cli -m gpt-oss-20b-GGUF-MXFP4.gguf --jinja -p "Explain MoE in one sentence." -st - ์ถ๋ก ๊ฐ๋: ์์คํ
ํ๋กฌํํธ์
Reasoning: low|medium|high์ง์ - Ollama:
ollama run hf.co/axteam/gpt-oss-20b-GGUF - KV ์บ์ ์ ๊ฐ(์ ํ):
--cache-type-k q8_0 --cache-type-v q8_0 -fa onโ KV ์บ์ ์ฝ ์ ๋ฐ, GGUF ํ์ผ๊ณผ ์๊ด์์ด ์คํ ์ ์ง์
๋ผ์ด์ ์ค
- ๋ผ์ด์ ์ค: Apache-2.0 โ ์๋ณธ ๋ชจ๋ธ์ ๋ผ์ด์ ์ค๋ฅผ ๊ทธ๋๋ก ๋ฐ๋ฆ
- ๊ถ๋ฆฌ: ์๋ณธ ๊ฐ์ค์นยทํ ํฌ๋์ด์ ยทchat template์ OpenAI ์์ , ์ด ๋ฆฌํฌ๋ GGUF๋ก ๋ค์ ๋ฌถ์ ๊ฒ