GPT-OSS-120B Memory Requirements: Where Every Gigabyte Goes, From 8K to 131K Context

GPT-OSS-120B is an Apache 2.0-licensed mixture-of-experts model released in August 2025 whose 60.8 GiB MXFP4 checkpoint generally requires about 65 GB of memory plus operational headroom, despite activating only 5.1 billion of its 116.8 billion parameters per token. Its KV cache remains relatively modest, adding under 5 GB at the full 131,072-token context window, although multiple simultaneous conversations can substantially increase memory use. The model fits most directly on an 80 GB A100 or H100-class GPU, while llama.cpp can support smaller GPUs by keeping some expert layers in system RAM; high-memory Macs and 128 GB unified-memory systems are also presented as viable options. Reported performance ranges from roughly 30 tokens per second for an RTX 5090 supplemented by system memory to nearly 200 tokens per second on a 96 GB RTX Pro 6000. Users without suitable local hardware can instead access a comparable hosted Qwen3.5 122B A10B model through an OpenAI-compatible API. The smaller GPT-OSS-20B requires about 16 GB of memory and offers competitive coding performance at lower hardware cost, but GPT-OSS-120B delivers stronger results on knowledge-heavy benchmarks. Runtime choices including Ollama, LM Studio, llama.cpp, and vLLM vary mainly in whether they require the full model in GPU memory, can offload parts to RAM, or support multi-user server deployment.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.