DEV.co
Open-Source LLM · RedHatAI

gpt-oss-120b

gpt-oss-120b is a 120-billion-parameter open-weight language model from OpenAI, released under Apache 2.0 license. It uses mixture-of-experts (MoE) architecture with only 5.1B active parameters and MXFP4 quantization to fit on a single 80GB GPU (H100/MI300X). Designed for reasoning-heavy tasks, agentic workflows, and production use cases. Requires the 'harmony' response format for correct operation.

Source: HuggingFace — huggingface.co/RedHatAI/gpt-oss-120b
120.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
40.7k
Downloads (30d)

Key facts

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

FieldValue
DeveloperRedHatAI
Parameters120.4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads40.7k
Likes5
Last updated2026-05-18
SourceRedHatAI/gpt-oss-120b

What gpt-oss-120b is

120B-parameter MoE model with 5.1B active parameters. Post-trained with MXFP4 quantization of MoE weights. Trained on OpenAI's harmony response format (mandatory for inference). Supports configurable reasoning effort (low/medium/high), full chain-of-thought output, function calling, Python execution, and structured outputs. Validated on vLLM 0.10.1.1, Red Hat OpenShift AI 2.25, and Red Hat AI Inference Server 3.2.2. Context length unknown. Compatible with Transformers and PyTorch ecosystems.

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="RedHatAI/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 reasoning workloads

Complex problem-solving, multi-step reasoning, and analytical tasks that benefit from chain-of-thought transparency and configurable reasoning effort. Full reasoning output aids debugging and trust validation.

Agentic systems and automation

Agent-driven applications leveraging function calling, code execution, and autonomous decision-making. Native support for Python execution and structured outputs simplifies agent orchestration.

Single-GPU production deployments

Organizations needing 120B-scale reasoning on commodity hardware (H100/MI300X). MXFP4 quantization and MoE sparsity enable single 80GB GPU deployment without external inference clusters.

Running & fine-tuning it

Single 80GB GPU (NVIDIA H100, AMD MI300X recommended). VRAM estimate: ~80GB for full model with MXFP4 quantization of MoE weights and 8-bit precision on other parameters. Training/fine-tuning requirements not specified; inference-optimized specs provided only.

Model card states 'Fully customize models to your specific use case through parameter fine-tuning' but does not specify LoRA/QLoRA feasibility, training frameworks, or memory requirements for fine-tuning. Requires review of OpenAI's harmony repository or downstream documentation for adapter-based training details. Full parameter fine-tuning likely requires multi-GPU setup.

When to avoid it — and what to weigh

  • Strict latency SLAs under 50ms per token — MoE activation and reasoning workloads introduce per-token latency overhead. Lower-latency use cases should consider gpt-oss-20b (21B parameters) or smaller dense models.
  • Harmony format incompatible legacy systems — Model requires mandatory harmony response format application. Systems unable to modify prompt/response handling or using alternative chat templates may require custom integration work.
  • Resource-constrained environments under 80GB VRAM — Even with quantization, full model requires ~80GB GPU memory. Environments limited to <16GB VRAM should use gpt-oss-20b or dense smaller models instead.
  • Tasks requiring unknown context lengths — Model card does not specify maximum context window. Applications requiring specific sequence length guarantees should verify against OpenAI documentation or empirically test before deployment.

License & commercial use

Apache License 2.0 (permissive OSI license). No copyleft restrictions, no patent claims. Permits free use, modification, distribution, and commercial deployment without attribution requirement beyond license inclusion.

Apache 2.0 is an OSI-approved permissive license explicitly designed for commercial use without restrictions. Model card confirms 'Build freely without copyleft restrictions or patent risk—ideal for experimentation, customization, and commercial deployment.' Commercial use is clearly permitted. No gating, no evaluation requirements, no enterprise licensing needed.

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

Model card does not discuss security posture, red-teaming results, or safety measures. Full chain-of-thought output is 'not intended to be shown to end users'—indicates potential sensitive reasoning disclosure risk if exposed. Quantization (MXFP4) and sparsity may affect robustness behavior; requires evaluation before security-sensitive deployments. Requires review of OpenAI's safety documentation and arxiv paper (2508.10925) for threat modeling.

Alternatives to consider

gpt-oss-20b (OpenAI)

Smaller MoE variant (21B parameters, 3.6B active) for latency-sensitive or resource-constrained deployments. Fits in 16GB VRAM. Same license, harmony format, agentic capabilities. Trade-off: lower reasoning capacity.

Meta Llama 3.1 405B

Open-weight dense alternative (Llama 2 license non-OSI; gated). Larger capacity, different inference characteristics. No MoE sparsity; requires multi-GPU clusters. Strong community support and fine-tuning tools.

Mistral Large (7B/22B/123B)

Open-weight dense models (Mistral license; verify terms). Comparable or smaller scale. Better latency characteristics. Simpler fine-tuning ecosystem. No harmony format dependency.

Software development agency

Ship gpt-oss-120b with senior software developers

gpt-oss-120b fits a 120B-parameter reasoning engine on a single 80GB GPU without licensing overhead. Use the vLLM quickstart, OpenShift AI templates, or Transformers integration to spin up in minutes. Review the arxiv paper and OpenAI blog for safety and quantization details before production deployment.

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 this model commercially?
Yes. Apache 2.0 is an OSI-approved permissive license that explicitly permits commercial use, modification, and distribution without copyleft or patent restrictions. No evaluation, licensing fee, or enterprise agreement required.
What GPU hardware is required?
Minimum 80GB VRAM (NVIDIA H100, AMD MI300X recommended). Model uses MXFP4 quantization and MoE sparsity to fit single-GPU deployment. Smaller gpt-oss-20b (21B) runs on 16GB GPUs. Inference on CPUs only is not recommended per deployment examples.
Do I need to use the harmony response format?
Yes. Model was trained exclusively on OpenAI's harmony format and will not work correctly without it. You must apply harmony format manually via chat template or the openai-harmony package if using transformers' model.generate() directly.
What is the model's maximum context length?
Not specified in the model card. Requires verification against OpenAI documentation or empirical testing. Red Hat deployment examples reference --max-model-len 32768 in vLLM, suggesting 32K context support, but this is inferred and should be confirmed.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If gpt-oss-120b is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to deploy advanced reasoning at scale?

gpt-oss-120b fits a 120B-parameter reasoning engine on a single 80GB GPU without licensing overhead. Use the vLLM quickstart, OpenShift AI templates, or Transformers integration to spin up in minutes. Review the arxiv paper and OpenAI blog for safety and quantization details before production deployment.