v2.2 · 2026-06-12

GPT-OSS 20B

gpt-oss-20b

REASONING

GPT-OSS 20B is a 21B reasoning model profile for local AI planning. This page records provider, license, context, quantization, hardware-fit estimates, setup hints, source links, and caveats so readers do not choose a model by name alone.

PROVIDER

OpenAI

FAMILY

GPT-OSS

MODEL TYPE

reasoning

PARAMETERS

21B (3.6B active)

MODALITIES

text

ARCHITECTURE

MoE Transformer; alternating dense and locally banded sparse attention; grouped multi-query attention; 24 layers; 32 experts; 4 active experts per token; MXFP4 MoE weights

CONTEXT WINDOW

131,072 tokens

TRAINING TOKENS

—

RELEASE DATE

2025-08-05

LICENSE & LINKS

Apache 2.0license text

MODEL CARD

huggingface.co/openai/gpt-oss-20b

HUGGING FACE

openai/gpt-oss-20b

GGUF REPOSITORIES

lmstudio-community/gpt-oss-20b-GGUF community conversion

Community GGUF repository referenced by the June 2026 model research.

bartowski/openai_gpt-oss-20b-GGUF community conversion

Community GGUF repository referenced by the June 2026 model research.

SETUP HINTS

OLLAMA

ollama run gpt-oss:20b

LM STUDIO

Available according to the supplied June 2026 research; select a cited GGUF/MLX build and verify current app metadata before relying on hands-on setup guidance.

LLAMA.CPP

Use the cited GGUF repository (lmstudio-community/gpt-oss-20b-GGUF) with current llama.cpp-compatible tooling.

QUANTIZATION FILE SIZES (GGUF)

Q4_K_M

11.67 GB

Q5_K_M

11.73 GB

Q8_0

12.1 GB

Quantization note: Q4_K_M for first local attempts unless a cited runtime recommends a different default.

RAM / VRAM ESTIMATES

MIN RAM

12 GB

COMFORTABLE RAM

16 GB

MIN VRAM

12 GB

COMFORTABLE VRAM

16 GB

These are conservative local-inference estimates, not official hardware requirements.

HARDWARE FIT

CPU only (no GPU)

limited

8 GB RAM

16 GB RAM

comfortable

32 GB RAM

comfortable

8 GB VRAM

12 GB VRAM

limited

24 GB VRAM

comfortable

Apple Silicon (unified memory)

comfortable

Hardware fit values are conservative local-inference estimates based on GGUF size plus runtime and KV-cache overhead. Actual requirements depend on context length, quantization, runtime, GPU offload, and other running apps.

BEST FOR

·Local reasoning, coding experiments, tool calling, and privacy-minded users with 16GB-plus memory and patience today.

AVOID IF

·Avoid if your machine has under 16GB memory or you expect plug-and-play speed today locally.

CAVEATS

·This is the realistic GPT-OSS model for local users, but it is still a large model.

·Use a Harmony-compatible runtime such as Ollama, LM Studio, vLLM, Transformers chat templates, or another integration that handles GPT-OSS formatting correctly.

·Tool use depends on the app or server exposing tools; the model alone does not provide web browsing, Python, or function execution.

·Higher reasoning effort can improve difficult tasks but will usually slow responses.

·Full chain-of-thought may be available for debugging, but user-facing products should show concise reasoning summaries instead.

·LM Studio lists 12GB minimum system memory, but 16GB-plus VRAM or unified memory is the safer beginner recommendation.

·The 128k context window is an upper limit; long prompts can be slow and memory-heavy.

·Common Q4, Q5, and Q8 GGUF variants save little space because GPT-OSS MoE feed-forward weights are generally kept at MXFP4.

·Commercial-use status from the supplied research: permitted, subject to OpenAI gpt-oss usage policy

·Beginner summary from supplied research: GPT-OSS 20B is OpenAI’s smaller open-weight reasoning model for capable local chat, coding, and tool-style workflows. It fits high-end consumer machines, but beginners must use compatible apps and expect variable speed locally with 128k context.

SOURCE URLS