OpenAI gpt-oss Open-Weight Models: Full Guide 2026

OpenAI shipped the gpt-oss open-weight model family on June 9, 2026. The release includes three sizes: gpt-oss-8b, gpt-oss-20b-a3b, and gpt-oss-120b-a12b. All three share a 256,000 token context window and use the same tokenizer. OpenAI published the weights on Hugging Face and linked the release from its official website. The company calls the release a response to teams that need local control and predictable pricing. This gpt-oss open-weight guide covers the technical details, license, and setup costs. The smallest model is an 8B dense model. The two larger models are mixture of experts designs. Each model is free to download, fine-tune, and deploy under the OpenAI Open Weight License 1.0.

The release matters because it changes how developers can use OpenAI style models. Teams can run gpt-oss on their own hardware without API billing or rate limits. The 20B MoE model activates 3B parameters per token and fits in a single 24GB consumer GPU. That makes it practical for local agentic coding, document processing, and offline use. The 120B MoE model activates 12B parameters and targets servers with multiple 80GB GPUs. Closed models from OpenAI still hold quality leads on some tasks, but the gpt-oss family closes the gap in coding, math, and instruction following. For context on other models, see best open-source LLM models in 2026.

The license is the most important part. OpenAI did not choose a standard Apache 2.0 or MIT license. The OpenAI Open Weight License 1.0 allows commercial use, modification, and redistribution of model weights. It does not include the training data or full training code. That makes the release open-weight, not open-source under the OSI definition. Some companies will accept the terms because they avoid API fees and data sharing. Others may prefer a standard permissive license from Qwen or Llama families. Benchmark details below show what the release can and cannot replace. This matters at a time when AI free tier limits are getting tougher.

How Do the Top Options Compare?

Benchmarks are vendor reported as of June 2026. Actual quality depends on quantization, sampling, and task. Parameters / Active refers to total parameters and active parameters per token for mixture-of-experts models.

1. gpt-oss-8b , Best for CPU and edge local inference

gpt-oss-8b is the entry point for the gpt-oss family. It is a dense 8 billion parameter model with a 256k token context window. It runs on CPU-only servers with enough RAM, Apple Silicon Macs, and low-end GPUs. At 4-bit quantization the model needs about 5GB of memory. At 8-bit it needs roughly 8GB. That makes it one of the few open-weight models that supports a 256k context on modest hardware. The small footprint helps edge deployments and air-gapped workstations.

OpenAI reports that gpt-oss-8b scores 61.2 percent on MMLU-Pro and 68.4 percent on HumanEval. Those results trail the 20B and 120B models, but they are respectable for an 8B dense model. The model handles long document processing, basic coding, classification, and local retrieval tasks. It will not match frontier closed models on hard reasoning or long-horizon agent work. Users should treat it as a local utility model rather than a flagship replacement.

For deployment, the weights work with standard local inference tools. You can run gpt-oss-8b in llama.cpp, Hugging Face Transformers, and Ollama after conversion. It supports 4-bit, 5-bit, and 8-bit quantization. On Apple Silicon with 16GB of RAM, users can run it with 8-bit or lower precision. CPU-only servers can handle small batch inference, but token speeds will fall well below GPU speeds.

For a wider list of self-hosted models, see top open-source LLMs to self-host free in 2026. The gpt-oss-8b weights are available on Hugging Face and linked from OpenAI. The OpenAI Open Weight License allows commercial use and fine-tuning, but it does not include the training data.

Key strengths:

• ✅ Runs on CPU-only machines and Apple Silicon with low memory overhead.
• ✅ Supports the full 256k context window even in the small dense model.
• ✅ Free for commercial use and fine-tuning under the OpenAI Open Weight License.
• ✅ Good base for offline document processing and embedded local tasks.
• ✅ Small enough for air-gapped environments with no API calls.
• ❌ Falls behind the 20B and 120B gpt-oss models on complex reasoning benchmarks.
• ❌ Dense 8B model has slower throughput than small quantized models in some CPU-only setups.
• ❌ OpenAI Open Weight License is not OSI approved and excludes training data.

Who it’s for: Choose the 8B model if you need local inference on a laptop, CPU server, or edge device with limited memory.

2. gpt-oss-20b-a3b , Best for single-GPU local coding and agents

gpt-oss-20b-a3b is the most practical option for developers and small teams. It is a mixture of experts model with 20 billion total parameters and 3 billion active parameters per token. Because only a fraction of the model activates per token, inference is faster than a dense 20B model. A single 24GB consumer GPU such as an RTX 4090 or 3090 can run it at 4-bit or 8-bit precision. That hardware profile removes the need for cloud API billing and data egress.

OpenAI reports benchmark scores of 78.9 percent on MMLU-Pro, 82.3 percent on HumanEval, and 58.1 percent on GPQA-Diamond. Those results place the 20B model close to some closed mid-tier APIs while running fully local. It handles agentic coding loops, tool calls, long document summarization, and structured output. It is not a drop-in replacement for OpenAI’s largest frontier models on hard research math or million-token agent state. But it covers a large share of daily coding and business tasks.

Deployment is easier than the 120B model. Many users run it through vLLM, Hugging Face Text Generation Inference, or llama.cpp with 4-bit quantization. The model file size is around 11GB at 4-bit and 20GB at 8-bit. A 24GB GPU fits 8-bit with limited KV cache, so teams often choose 4-bit or 5-bit for longer context. The full 256k context is supported, but longer prompts consume more VRAM.

Local deployment changes cost math. Teams can pay once for hardware instead of tracking token spend. For more on free and low-cost options, see best free AI models in 2026 with no API costs. The weights are linked from the OpenAI release page and mirrored on Hugging Face. You can also inspect community serving guides on GitHub.

Key strengths:

• ✅ Fits on one 24GB GPU at 4-bit or 8-bit precision, so no multi-GPU setup is needed.
• ✅ 3B active parameters per token deliver fast inference compared with dense 20B models.
• ✅ Strong coding and tool-use benchmark scores for a local model.
• ✅ Commercial use and fine-tuning are allowed under the OpenAI Open Weight License.
• ✅ Full 256k context supports long documents and long agent sessions.
• ❌ Requires a modern GPU with at least 24GB VRAM for comfortable inference.
• ❌ MoE model requires more total storage than a dense model of similar active size.
• ❌ Open-weight license limits visibility into training data and full training code.

Who it’s for: Choose the 20B model if you want strong local coding agents and document work on a single 24GB GPU without API fees.

3. gpt-oss-120b-a12b , Best for data center self-hosting and frontier-adjacent quality

gpt-oss-120b-a12b is the flagship of the gpt-oss release. It has 120 billion total parameters and 12 billion active parameters per token. It targets teams that own or rent multi-GPU servers, usually two or more 80GB GPUs at 8-bit precision. It can run at lower quantization on two 48GB GPUs, but throughput and quality drop. The model uses the same 256k tokenizer as the smaller gpt-oss models. That means teams can test prompts and tools on the 8B or 20B model and then scale up to the 120B model on server hardware.

OpenAI reports strong results on broad benchmarks. The company lists 84.7 percent on MMLU-Pro, 86.9 percent on HumanEval, and 64.8 percent on GPQA-Diamond. Those scores sit closer to closed frontier systems than the smaller gpt-oss models. The 120B model also scores 91.2 percent on MATH-500 and handles complex agentic coding with multiple file edits, tool calls, and long context. It is still not a copy of OpenAI’s closed frontier model, and the company does not claim parity on every task.

Operational cost is the main downside. Serving a 120B MoE model requires real GPU capacity, power, and cooling. At 8-bit precision the weights need about 120GB of VRAM, which usually means two 80GB GPUs with careful offloading. At 4-bit precision the model needs about 60GB, but quality falls on hard reasoning. Batch size, active expert routing, and KV cache all change memory use. Teams should benchmark their own workloads before committing to reserved cloud instances or hardware purchases.

Many teams will use the 20B model for interactive work and only call the 120B model for batch jobs or hard problems. The 120B model also benefits from continuous batching in vLLM and SGLang. For broader pricing context, see major AI API pricing model updates June 2026. The weights are available on Hugging Face and documented on the OpenAI site.

Key strengths:

• ✅ Best benchmark scores in the gpt-oss family, close to closed frontier quality.
• ✅ 12B active parameters per token make serving efficient relative to dense 120B models.
• ✅ Same tokenizer and 256k context as smaller models, so prompt and tooling transfer easily.
• ✅ Commercial use and fine-tuning are allowed under the OpenAI Open Weight License.
• ✅ Strong coding, math, and instruction-following results for an open-weight model.
• ❌ Requires multi-GPU infrastructure, usually two or more 80GB GPUs for comfortable 8-bit serving.
• ❌ High serving cost and power draw compared with the 20B and 8B models.
• ❌ OpenAI Open Weight License excludes training data and is not a standard OSI license.

Who it’s for: Choose the 120B model if you run a data center or GPU cluster and need the strongest gpt-oss quality for coding, agents, and large batch work.

Frequently Asked Questions

What Should You Remember?

• Open-weight release: OpenAI shipped gpt-oss-8b, gpt-oss-20b-a3b, and gpt-oss-120b-a12b on June 9, 2026.
• Three sizes: 8B dense, 20B MoE, and 120B MoE models share a 256k token context window.
• License: The OpenAI Open Weight License 1.0 allows commercial use and fine-tuning but excludes training data.
• Local hardware: The 20B model fits on one 24GB GPU, while the 120B model needs multi-GPU servers.
• Benchmarks: The 120B model posts 84.7 percent MMLU-Pro and 86.9 percent HumanEval, near closed frontier levels.
• Cost control: Self-hosting removes per-token API fees but shifts cost to GPUs, power, and operations.

Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.