Best larger-system route
Team 32
The smallest system in the range that reaches this model's preferred working allowance.
Explore Team 32OpenAI model
GPT-OSS 20B Hardware Requirements & Server Compatibility
GPT-OSS 20B is a current OpenAI release for lower-latency reasoning, tool use and local specialised assistants. The pinned repository contains approximately 41.274GB of model weights. Our 16GB on one GPU figure is a vendor documented memory screen, while 32GB on one GPU is the safer starting allowance for deployment testing.
- Model version
openai/gpt-oss-20b- Family and variant
- GPT-OSS · 20B MXFP4
- Source version
6cee5e81ee83- Source updated
- 26 August 2025
OpenAI GPT-OSS 20B
openai/gpt-oss-20b - Minimum GPU memory
- 16GB on one GPU
- Recommended hardware
- 32GB on one GPU
- Licence
- Apache License 2.0
- Useful for
- Language & reasoning, Coding & agents
Buyer verdict
Where GPT-OSS 20B is a sensible fit
Shortlist GPT-OSS 20B when lower-latency reasoning, tool use and local specialised assistants is the priority and the exact licence and runtime suit the organisation. Choose hardware from the recommended allowance, then measure quality and performance on representative work before purchase.
These figures apply to the named model version. Quantisation, fine-tuning, context length, image resolution, batch size and serving software can materially change the hardware needed.
Hardware requirements
- Minimum
- 16GB on one GPU
- Recommended
- 32GB on one GPU
This is the developer's published memory target. Usable context and throughput remain workload-dependent.
This working allowance creates room for runtime allocations and representative workload testing; it is not a performance benchmark.
Technical specification
GPT-OSS 20B model and hardware facts
Specifications shown for source version
6cee5e81ee83917806bbde320786a8fb61efebee, updated
26 August 2025.
- Architecture
- Sparse mixture of experts
- Parameters
- 21B total / 3.6B active
- Context
- 131,072 tokens Architecture ceiling; serving capacity requires testing
- Modalities
- Text → text
- Repository weights
- 41.274GB Pinned first-party repository file total
- Native format
- MXFP4 MoE weights
Product compatibility
GPT-OSS 20B compatibility across all 11 GPU systems
- Systems that fit without model splitting
- 11
- Systems needing multi-GPU validation
- 0
A system is listed as fitting when its GPU memory meets the requirement shown above. Speed, usable context, batch size and concurrent users still need testing with the final model and software configuration.
sme workstation
Team 32
- Per GPU
- 32GB
- Total GPU memory
- 32GB
- GPU count
- 1
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
sme workstation
Company 64
- Per GPU
- 32GB
- Total GPU memory
- 64GB
- GPU count
- 2
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
sme workstation
Studio 96
- Per GPU
- 96GB
- Total GPU memory
- 96GB
- GPU count
- 1
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
sme workstation
Studio 192
- Per GPU
- 96GB
- Total GPU memory
- 192GB
- GPU count
- 2
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
pcie rack
Value Rack 128
- Per GPU
- 32GB
- Total GPU memory
- 128GB
- GPU count
- 4
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
pcie rack
Value Rack 256
- Per GPU
- 32GB
- Total GPU memory
- 256GB
- GPU count
- 8
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
pcie rack
Enterprise 384
- Per GPU
- 96GB
- Total GPU memory
- 384GB
- GPU count
- 4
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
pcie rack
Enterprise 768
- Per GPU
- 96GB
- Total GPU memory
- 768GB
- GPU count
- 8
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
pcie rack
H200 1.1TB
- Per GPU
- 141GB
- Total GPU memory
- 1,128GB
- GPU count
- 8
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
frontier partner
Frontier Native 2.3TB
- Per GPU
- 288GB
- Total GPU memory
- 2,304GB
- GPU count
- 8
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
frontier partner
Frontier Rack 20TB
- Per GPU
- 288GB
- Total GPU memory
- 20,736GB
- GPU count
- 72
Also meets this model page's recommended working allowance.
Available GPU memory exceeds the model developer's published requirement. Confirm speed, context, batch size and concurrent users before purchase.
Deployment reality
Strengths, limits and runtime route
What it is good at
- Officially described as running within 16GB memory.
- Native configurable reasoning effort and tool use.
- Apache 2.0 licensing and several documented local runtimes.
- Smaller active parameter count than the 120B sibling for lower-latency work.
Where to be cautious
- The Harmony response format must be preserved by the serving stack.
- The repository contains alternative files, so its total download size is not the same as the published loaded-memory target.
- Repository weight size does not include every runtime allocation, KV cache, media encoder, batch or concurrent session.
- No GPU Servers benchmark result is claimed until this exact revision has been run under disclosed test conditions.
Serving software
- vLLM
- Transformers
- Ollama
- llama.cpp
Before installation
- Use source version 6cee5e81ee83917806bbde320786a8fb61efebee rather than an unversioned latest branch.
- Start with vLLM only after checking support for the exact architecture and numerical format.
- Use 32GB on one GPU as the procurement starting point; the lower figure is a minimum-memory screen, not a service-level promise.
- Record precision, context, batch, concurrency, container digest, driver and measured latency in the acceptance report.
System requirements
- GPU layout
- The minimum is screened against one GPU. Additional GPUs provide replicas or throughput unless the chosen runtime supports model parallelism.
- Context and cache
- 131,072 tokens. Context length, KV cache, batch and concurrent sessions must be tested together; the architecture limit is not a throughput guarantee.
- System RAM
- Plan system RAM above the 41.274GB repository-weight footprint where model staging or CPU offload is required.
- Storage
- Reserve at least 104GB for the pinned weights, runtime cache and one rollback copy; production datasets and logs are additional.
- Serving software
- First-party material names vLLM, Transformers, Ollama, llama.cpp. Pin the serving version and container digest because support for recent architectures can change quickly.
- Representative workload
- Acceptance-test lower-latency reasoning, tool use and local specialised assistants at the intended quality, context, batch, concurrency and response-time target.
Performance: speed, usable context and concurrency depend on the selected system, software and workload. No benchmark is quoted on this page.
Commercial and legal boundary
Apache License 2.0
Commercial use: permitted
OpenAI releases the model under Apache 2.0, permitting commercial use subject to the licence terms and notices.
Always retain the applicable notices and recheck the live terms for the intended organisation, territory, use and distribution route. Obtain legal advice where required; the official licence governs use.
Read the official licenceTechnical questions
GPT-OSS 20B deployment FAQ
What GPU memory does GPT-OSS 20B need?
Use 16GB on one GPU as the lower memory screen and 32GB on one GPU as the safer starting allowance. The final requirement changes with runtime, context, batch, concurrency and precision.
Which GPU Servers product should I start with for GPT-OSS 20B?
Use the compatibility table to find systems that meet the recommended allowance. A system that only meets the minimum can load the model in principle but may not meet the required context or response time.
Can GPT-OSS 20B be used commercially?
OpenAI releases the model under Apache 2.0, permitting commercial use subject to the licence terms and notices. The linked official licence is authoritative; legal advice may be appropriate for the intended use and distribution route.
What a complete GPT-OSS 20B deployment needs
GPU memory is only one part of the system. Storage, data access, serving software, monitoring and administrator handover also affect a reliable deployment.
