About GPT-OSS 120B
GPT-OSS 120B is the larger of OpenAI's pair of Apache 2.0 open-weight releases, aimed at production-grade reasoning on a single high-memory accelerator rather than a cluster. It is a mixture-of-experts model in which only a few billion parameters activate per token — the source of its speed — while the full expert set defines its memory footprint.
OpenAI positions it near its proprietary mini-tier reasoning models on capability, with the differences that define open weights: inspectable chain-of-thought, local fine-tuning, and freedom to deploy wherever the hardware exists.
GPT-OSS 120B is OpenAI's 117B-parameter model released in August 2025, with a 128K-token context window. It is a mixture-of-experts model: all 117B parameters must sit in memory, but only ~5.1B 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 0.3 GB at an 8K context, 4.8 GB at 128K, and 4.8 GB at the full 128K window (FP16 cache).
The MXFP4-native release defines the planning number: OpenAI sized GPT-OSS 120B to fit a single datacenter-class accelerator, and the quantisation table above shows why nothing much smaller gets close — every expert must stay resident, so the small active-parameter count buys speed, not memory relief. In practice that means high-memory workstation cards, unified-memory machines at the top of Apple's range, or small multi-GPU rigs. The hybrid attention layout keeps long-context KV growth gentle, so once the weights fit, context length is rarely the constraint. The runtime ecosystem matches its smaller sibling's.
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 120B need?
At MXFP4 (native) with an 8K context, GPT-OSS 120B needs about 74 GB (weights 66 GB + KV cache + overhead). The smallest common hardware that fits is a A100 80GB.
Can an RTX 4090 (24GB) run GPT-OSS 120B?
Not fully in VRAM. GPT-OSS 120B needs about 74 GB even at MXFP4 (native), so a 24 GB card would have to offload layers to system RAM at a large speed penalty.
Can a Mac run GPT-OSS 120B?
Yes — Apple Silicon with 128 GB of unified memory or more (macOS lets the GPU use ~75% of it, ~96 GB) runs GPT-OSS 120B at MXFP4 (native).
How much VRAM does GPT-OSS 120B need at its full 128K context?
The KV cache grows with context: about 0.3 GB at 8K tokens versus 4.8 GB at the full 128K window (FP16 cache). Add that to the weights (66 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: Mistral Small 4 (119B) · Llama 4 Scout (109B) · Llama 3.3 70B (70.6B) · DeepSeek V4 Flash (284B) · Qwen3.6 35B-A3B (35B)
Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.