What GPU do I need to run openai/gpt-oss-20b?
A 21B (MoE) reasoning model whose MXFP4 weights run within 16GB. 20.9B parameters, published in MXFP4. View on Hugging Face
What gpt-oss-20b is
gpt-oss-20b is a 21B-parameter mixture-of-experts language model published by OpenAI on Hugging Face, the smaller of two open-weight gpt-oss models aimed at lower-latency, local and specialized use cases. Its MoE weights are post-trained in MXFP4 precision, which the model card says is what lets it run within 16GB of memory. It supports configurable reasoning effort (low, medium, high), function calling, web browsing and Python code execution, and is released under Apache 2.0.
License note: fully permissive, no gating and no usage restrictions. Facts in this section are sourced from gpt-oss-20b's Hugging Face model card, not benchmarked by Aquanode.
What it's used for
- Low-latency and local/on-device inference
- Reasoning and chain-of-thought tasks
- Agentic tool use and function calling
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| MXFP4 (native) | 12.8 GB | 16.0 GB | RTX 5060 Ti | 1 | $0.165/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
OpenAI's own model card: MXFP4-quantized MoE weights let gpt-oss-20b run within 16GB of memory.
Cheapest way to run gpt-oss-20b at its published (MXFP4) precision: 1× RTX 5060 Ti, at $0.165/hr per GPU ($0.165/hr total).
gpt-oss-20b: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Sliding-window attention: windowed layers stop growing at the window, the full-attention layers keep growing. Cached per token: 49,152 bytes at 16-bit while the context is inside the window.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.10 GB |
| 32K | 0.75 GB |
| 128K | 3.0 GB |
Computed from the layer, head and window counts in the model's own configuration (openai/gpt-oss-20b), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
How to run gpt-oss-20b
Run gpt-oss-20b with vLLM
From openai/gpt-oss-20b's own model card (requires vllm>=0.10.1+gptoss).
vllm serve openai/gpt-oss-20bSource: https://huggingface.co/openai/gpt-oss-20b/raw/main/README.md
Run gpt-oss-20b with Ollama
From the model card's own Ollama instructions; verified against Ollama's own library listing.
ollama run gpt-oss:20bRun gpt-oss-20b with GGUF quantizations
Prebuilt GGUF weights published at unsloth/gpt-oss-20b-GGUF. Run with llama.cpp's llama-server or load the repo directly in LM Studio.
llama-server -hf unsloth/gpt-oss-20b-GGUFDeploy gpt-oss-20b on Aquanode
Aquanode has no one-click deploy template for gpt-oss-20b; you install the inference engine yourself with the commands below. Aquanode sells GPU pods billed per second, not a hosted inference API.
- Launch a bare GPU pod sized to the requirement above (1× RTX 5060 Ti or larger).
- Open a terminal on the pod, or save one of the commands above as a startup script so it runs automatically the first time the pod boots.
- Run the command and connect to the resulting endpoint.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More gpt-oss models
- gpt-oss-20b-BF16 (20.9B, BF16)
- Huihui-gpt-oss-20b-BF16-abliterated (20.9B, BF16)
- gpt-oss-120b (116.8B, MXFP4)
- gpt-oss-120b-Eagle3-short-context (787M, BF16)
- EAGLE3-gpt-oss-20b (350M, BF16)
Lineage
Quantized and fine-tuned versions
Alternatives at this size
Other models for text-generation within about a third of gpt-oss-20b's 20.9B parameters, from other model lines.
- Qwen2.5-14B-Instruct (14.8B, BF16)
- GLM-4.7-Flash (31.2B, BF16)
- gpt-neox-20b (20.7B, F16)
- DeepSeek-Coder-V2-Lite-Instruct (15.7B, BF16)
- phi-4 (14.7B, BF16)
More on gpt-oss-20b
Fits on a 16 GB GPU at MXFP4: every model that fits in 16 GB.
Best reasoning models: how gpt-oss-20b ranks against the rest.
Best chat and assistants models: how gpt-oss-20b ranks against the rest.
Related reading: H100 pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
Related on Aquanode: GPU requirements for idefics-9b, GPU requirements for Olmo-Hybrid-7B, GPU requirements for Agents-A1-4B, and GPU requirements for XCurOS0.1-8B-Instruct.
Your own GPU cloud accounts
One console for the GPU clouds you already rent from. Your keys stay on your machine.
$20 a month, flat