DEV.co
Open-Source LLM · openai

gpt-oss-120b

gpt-oss-120b is OpenAI's 120-billion-parameter open-weight model released under Apache 2.0. It uses mixture-of-experts (MoE) with 5.1B active parameters, runs on a single 80GB GPU via MXFP4 quantization, and is designed for reasoning-heavy, agentic tasks. The model requires the proprietary 'harmony' response format and supports tool use (function calling, web browsing, code execution). It is production-ready, fully fine-tunable, and commercially usable under Apache 2.0.

Source: HuggingFace — huggingface.co/openai/gpt-oss-120b
120.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
4.3M
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developeropenai
Parameters120.4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads4.3M
Likes5k
Last updated2025-08-26
Sourceopenai/gpt-oss-120b

What gpt-oss-120b is

gpt-oss-120b is a 117B-parameter MoE transformer (5.1B active) trained with OpenAI's harmony response format. It quantizes to MXFP4 and fits in 80GB GPU memory (H100, MI300X). Inference frameworks include Transformers (with chat template), vLLM (custom gpt-oss wheel), Ollama, LM Studio, and PyTorch/Triton reference implementations. The model exposes full chain-of-thought reasoning and supports configurable reasoning effort (low/medium/high) via system prompts. Context length is unknown. Fine-tuning is supported on single H100 nodes.

Quickstart

Run gpt-oss-120b locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="openai/gpt-oss-120b")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.

Deployment

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

Production Agentic AI Applications

Deploy as a self-hosted reasoning engine for autonomous agent workflows requiring function calling, web browsing, and code execution without reliance on external APIs.

Enterprise RAG and Custom LLM Apps

Fine-tune on proprietary domain data (e.g., technical docs, legal contracts) and self-host for full data privacy and compliance control.

Reasoning-Heavy Problem Solving

Leverage exposed chain-of-thought for tasks like technical analysis, debugging, mathematical reasoning, or complex multi-step planning where model reasoning transparency is valuable.

Running & fine-tuning it

Minimum: Single 80GB GPU (NVIDIA H100, AMD MI300X) for full MXFP4-quantized inference. Estimate: ~80GB VRAM for inference (verify with your framework). Fine-tuning: single H100 node (exact memory/batch size unknown; requires testing). Smaller gpt-oss-20b variant runs in 16GB.

Full parameter fine-tuning is supported and documented. No LoRA/QLoRA guidance provided in the card. Fine-tuning on single H100 is feasible for the 120B model. Start with lower learning rates and test on a held-out validation set due to the MoE architecture. Harmony format compliance must be maintained post-tuning.

When to avoid it — and what to weigh

  • Latency-Critical Consumer Applications — The 120B model is inference-heavy; the smaller gpt-oss-20b is recommended for sub-second response requirements. Reasoning effort 'high' will add significant latency.
  • Limited Hardware Infrastructure — Requires 80GB GPU (H100 or MI300X) minimum. Not viable for edge devices or servers with <40GB VRAM without further quantization experimentation.
  • Strict Context-Length Constraints — Context length is not specified in the card. Verify against your use case requirements before committing to production deployment.
  • Non-Harmony Format Workflows — The model is specifically trained on harmony format and will not work correctly without it. Retrofitting existing prompt templates is required.

License & commercial use

Apache License 2.0 (Apache-2.0). This is a permissive OSI-approved license with no copyleft, patent indemnity, and no commercial use restrictions.

Commercial use is explicitly permitted under Apache 2.0. The card states: 'Permissive Apache 2.0 license: Build freely without copyleft restrictions or patent risk—ideal for experimentation, customization, and commercial deployment.' You may use, modify, and deploy this model commercially provided you include the license notice and disclaimer in distributions.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

The card does not discuss security evaluations, adversarial robustness, or known vulnerabilities. Chain-of-thought exposure is a design choice for transparency but may leak intermediate reasoning in multi-user environments—consider access controls. MXFP4 quantization may have undocumented numerical stability trade-offs. Self-hosting eliminates API-level data transmission risk but requires infrastructure hardening. No explicit safety filter or content policy is documented.

Alternatives to consider

gpt-oss-20b (OpenAI)

Same architecture and license, but 20B parameters (3.6B active) with lower latency for local/edge deployment or lower hardware costs. Choose this if <80GB GPU availability or <500ms response times are mandatory.

Llama 3.3-70B or Llama 2 (Meta)

Also open-weight and permissively licensed (Llama Community License 2.1), larger parameter count (70B–405B). Llama 3.3 lacks MoE but may have broader community tooling; evaluate performance on your benchmarks.

Mixtral 8x22B (Mistral AI)

Open MoE model with permissive license (Apache 2.0). Smaller total size (176B) with sparser active compute (39B active), but lacks OpenAI's reasoning-tuning and harmony format. Lower hardware barriers for some workloads.

Software development agency

Ship gpt-oss-120b with senior software developers

Start with vLLM, Transformers, or Ollama inference. If your team needs architecture guidance, fine-tuning support, or private deployment help, we can connect you with specialists who've optimized MoE models at scale.

Talk to DEV.co

Related 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-120b FAQ

Can I use gpt-oss-120b commercially without paying OpenAI?
Yes. The Apache 2.0 license permits commercial use, modification, and distribution. You are not required to pay licensing fees to OpenAI. You must include the Apache 2.0 license text and disclaimer in your distribution.
What GPU do I need to run this model?
You need a single GPU with at least 80GB VRAM, such as NVIDIA H100 or AMD MI300X, to run the full MXFP4-quantized model. Smaller quantizations (e.g., INT4) or the 20B variant may fit in smaller GPUs, but are not officially documented.
Do I have to use the 'harmony' response format?
Yes. The card explicitly states the model is 'trained on harmony response format and should only be used with the harmony format as it will not work correctly otherwise.' Transformers pipelines apply it automatically; manual inference requires the openai-harmony package.
What is the model's context length?
Not specified in the card or model metadata. Check the arXiv paper (2508.10925) or contact OpenAI directly for this critical inference parameter.

Custom software development services

Need help beyond evaluating gpt-oss-120b? 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-120b?

Start with vLLM, Transformers, or Ollama inference. If your team needs architecture guidance, fine-tuning support, or private deployment help, we can connect you with specialists who've optimized MoE models at scale.