gpt-oss-20b
gpt-oss-20b is OpenAI's 21B-parameter open-weight language model designed for lower-latency and local deployment. It uses a mixture-of-experts architecture with 3.6B active parameters, supports agentic capabilities (function calling, web browsing, code execution), and is governed by a permissive Apache 2.0 license. The model requires harmony response format and can run on 16GB VRAM with MXFP4 quantization.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | openai |
| Parameters | 21.5B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 7M |
| Likes | 4.8k |
| Last updated | 2025-08-26 |
| Source | openai/gpt-oss-20b |
What gpt-oss-20b is
A 21.5B-parameter sparse MoE transformer released by OpenAI in August 2025. Architecture features MXFP4-quantized weights, supports configurable reasoning effort (low/medium/high), full chain-of-thought access, and agentic tool use. Inference support across Transformers, vLLM, PyTorch/Triton, Ollama, and LM Studio. Model card references arXiv:2508.10925 but context window length is not specified in provided data.
Run gpt-oss-20b locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="openai/gpt-oss-20b")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
ESTIMATE: ~16 GB VRAM (with MXFP4 quantization). Model card states gpt-oss-20b runs within 16GB of memory; MXFP4 quantizes MoE weights. VRAM requirements will be higher with full precision (FP32/FP16 likely requires 40+ GB). Verify against your target inference framework and batch size.
Full parameter fine-tuning is supported and documented as feasible on consumer hardware (single GPU). No mention of LoRA or QLoRA in the model card. For resource-constrained fine-tuning, assess LoRA compatibility independently; harmony format requirement may apply to fine-tuned outputs.
When to avoid it — and what to weigh
- Extremely Low Latency is Non-negotiable — At 21B active parameters, response time will be slower than smaller models (7B–13B). If sub-100ms inference is critical, evaluate smaller quantized alternatives.
- Production Context Window Unknown — Context length is not stated in the model card. If multi-turn conversation or document-in-context reasoning is central to your use case, verify context window capabilities before committing.
- Strict Copyleft or Patent Indemnity Requirements — While Apache 2.0 is permissive, it does not include explicit patent indemnity. If your legal team mandates broad patent protection (e.g., for derivative works), consult legal counsel.
- No Multi-modal Support — This is a text-only model; it does not process images, audio, or other modalities.
License & commercial use
Apache License 2.0 (SPDX: apache-2.0). Permissive OSI-compliant license allowing commercial use, modification, and distribution without copyleft obligations or forced publication of derivatives.
Apache 2.0 is a permissive license with no restrictions on commercial deployment. Model weights are open (not gated). However: (1) Apache 2.0 includes no patent indemnification clause—review patent exposure with legal counsel if patent risk is a concern; (2) model card does not discuss SLA, liability, or support; (3) ensure compliance with any fine-tuned model licensing if used in a commercial product. No commercial use restrictions are stated.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Strong |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Strong |
| Assessment confidence | High |
Model card does not discuss adversarial robustness, jailbreak resilience, or content filtering. Chain-of-thought access is noted as not intended for end-user display (implies reasoning artifacts may reveal internal logic). No security audit, red-teaming results, or mitigations are documented. Recommend threat-modeling for agentic use cases (code execution, web browsing) before production deployment. MXFP4 quantization may affect model behavior; test edge cases in your domain.
Alternatives to consider
gpt-oss-120b
Same family, larger (120B) and higher reasoning capability; requires 80GB GPU (H100 or MI300X). Choose if latency/hardware budget permits and reasoning depth is priority.
Llama 3.1 (70B or smaller variants)
Meta's open model with broader ecosystem support and eval coverage; non-OSI license (LLAMA 3.1 Community License)—requires explicit review for commercial use.
Mistral 7B or Mixtral
Smaller/faster alternatives with proven agentic support and well-documented fine-tuning. Trade reasoning depth for latency and resource efficiency.
Ship gpt-oss-20b with senior software developers
Download the model from Hugging Face Hub, verify context window and quantization strategy for your hardware, and test harmony format compliance in a dev environment before production rollout. Review patent and liability considerations with legal counsel.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
gpt-oss-20b FAQ
Can I use gpt-oss-20b in a commercial product?
What GPU memory do I need to run gpt-oss-20b?
Does gpt-oss-20b work without the harmony response format?
What is the context window length?
Software development & web development with DEV.co
Need help beyond evaluating gpt-oss-20b? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.
Ready to Deploy gpt-oss-20b?
Download the model from Hugging Face Hub, verify context window and quantization strategy for your hardware, and test harmony format compliance in a dev environment before production rollout. Review patent and liability considerations with legal counsel.