About GPT-OSS 20B
GPT-OSS 20B is one of the two open-weight models OpenAI released under the Apache 2.0 licence — its first open-weight language models since the early GPT-2 era. The smaller of the pair targets consumer hardware and on-device use: a mixture-of-experts design whose active-parameter count per token is a small fraction of the total, tuned for reasoning and tool use in agentic setups.
OpenAI ships it with its harmony chat format and visible chain-of-thought, and the genuinely open licence means it can be fine-tuned, redistributed and deployed commercially without a usage agreement.
GPT-OSS 20B is OpenAI's 21B-parameter model released in August 2025, with a 128K-token context window. It is a mixture-of-experts model: all 21B parameters must sit in memory, but only ~3.6B are active per token, which is what makes it fast for its size. It uses a hybrid-attention design — only a fraction of its layers cache the full context, so long conversations cost far less VRAM than a classic dense model: its KV cache is about 0.2 GB at an 8K context, 3.2 GB at 128K, and 3.2 GB at the full 128K window (FP16 cache).
OpenAI publishes GPT-OSS 20B already quantised in MXFP4, so the shipped weights are the size that matters — there is no meaningful full-precision download to plan around. That native size is what puts it in reach of gaming-class GPUs and unified-memory laptops, exactly the deployment OpenAI describes for it. Its attention design alternates sliding-window and full layers, keeping the KV cache small even deep into the context window, which the totals above reflect. Support landed in llama.cpp, Ollama and LM Studio at release, making it one of the simplest large models to run locally.
OpenAI ships the weights pre-quantised in MXFP4, so that is the size that matters — there is no meaningful FP16 variant to download. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.
Frequently asked questions
How much VRAM does GPT-OSS 20B need?
At MXFP4 (native) with an 8K context, GPT-OSS 20B needs about 14 GB (weights 12 GB + KV cache + overhead). The smallest common hardware that fits is a RTX 5060 Ti 16GB.
Can an RTX 4090 (24GB) run GPT-OSS 20B?
Yes. An RTX 4090's 24 GB runs GPT-OSS 20B at MXFP4 (native) (about 14 GB at 8K context).
Can a Mac run GPT-OSS 20B?
Yes — Apple Silicon with 24 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~18 GB) runs GPT-OSS 20B at MXFP4 (native).
How much VRAM does GPT-OSS 20B need at its full 128K context?
The KV cache grows with context: about 0.2 GB at 8K tokens versus 3.2 GB at the full 128K window (FP16 cache). Add that to the weights (12 GB at MXFP4 (native)) plus overhead — long contexts can cost more than a whole quantisation step.
Related
VRAM calculator for any model ·Token counter
Similar-size local models: Gemma 4 26B-A4B (26B) · Qwen3.6 27B (27B) · Phi-4 (14.7B) · Qwen3 Coder 30B-A3B (30.5B) · Gemma 4 31B (31B)
Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.