gpt-oss-20b

Name: gpt-oss-20b
Price: 0.02 USD
Author: OpenAI

OpenAI · gpt-oss · released 2025-08-05 · data checked 2026-11-01

gpt-oss-20b is a 20.9B-parameter Mixture-of-Experts model (3.6B active per token) from OpenAI (United States), released 2025-08-05 under the Apache-2.0 license. It scores 64.0 on the Small LM Index, #13 of 34 medium models. It needs about 13.8 GB of memory at 4-bit with an 8K-token context and costs $0.020 per million tokens through the cheapest API provider.

Runs in ~16 GB of memory thanks to native MXFP4 weights.

Small LM Index

64.0

#13 of 34 scored medium models

Likely 60.4–67.4 · from 6 of 6 core benchmarks

Parameters

20.9B

3.6B active · MoE

Price per 1M tokens

$0.020

$0.011 in / $0.045 out · blended 3:1

Context window

131K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best medium model on that benchmark. Sources: 4 reported by OpenAI, 3 measured by another lab. How we label sources.

MMLU-Pro Other lab Knowledge · Measured by NVIDIA75.0

Class best: Qwen3.6 27B 86.2

GPQA Diamond Lab-reported Knowledge71.5

Class best: Nanbeige4.3 33B 89.5

Humanity's Last Exam Lab-reported Reasoning10.9

Class best: Qwen3.8 27B 30.8

AIME Lab-reported Math · AIME 2025, no tools (high reasoning)91.7

Class best: Ternary Bonsai 2 27B 95.8

LiveCodeBench Other lab Coding · LiveCodeBench v6, measured by NVIDIA (Alibaba measured 74.6)61.0

Class best: Qwen3.8 27B 90.3

SWE-bench Verified Lab-reported Coding60.7

Class best: Qwen3.6 27B 77.2

IFBench Other lab Instruction following · Measured by NVIDIA65.0

Class best: Sage-A1.5 81.6

Specifications

Lab: OpenAI (US · United States)
Architecture: Mixture-of-Experts
Total parameters: 20.9B
Active parameters: 3.6B
Size class: Medium
Context window: 131K tokens
Input modalities: text
Reasoning: Reasoning (thinking)
License: Apache-2.0
Openness: Open weights
Memory to run: ~13.8 GB at 4-bit, 8K context (more sizes)
Hugging Face: 6M downloads (30 days) · 5.3K likes
API providers: 9 on OpenRouter
API price: $0.011 in / $0.045 out per 1M (OpenRouter, 2026-11-03)

Weights on Hugging Face ↗Try on OpenRouter ↗

Compare with

gpt-oss-20b vs Tri-21B Think (20.7B, index 43.6)
gpt-oss-20b vs EuroLLM 22B Instruct (22.6B, index 19.2)
gpt-oss-20b vs LFM2.5 24B-A2B (24B, index 61.4)
gpt-oss-20b vs Magistral Small 1.2 (24B, index 55.9)

Ranked in

#8 in Best small open models for math
#4 in Best LLMs that run on a 16 GB GPU
#6 in Best cheap LLM APIs (open models)
#7 in Best American open models under 100B

Sources

Something wrong here? Report it.

Run it locally

Uses ggml-org/gpt-oss-20b-GGUF (MXFP4). Swap the tag for another size: BF16, MXFP4, Q8_0.

Ollama

ollama run hf.co/ggml-org/gpt-oss-20b-GGUF:MXFP4

llama.cpp

llama-server -hf ggml-org/gpt-oss-20b-GGUF:MXFP4

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/gpt-oss-20b-OptiQ-4bit

vLLM (GPU server)

vllm serve openai/gpt-oss-20b

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	131K context
16-bit F16	43 GB	44 GB	46 GB
8-bit Q8_0	23 GB	24 GB	26 GB
4-bit Q4_K_M	14 GB	14 GB	17 GB

Check it against your GPU, Mac or phone →