OpenAI has released the weights for two models: gpt-oss-120b and gpt-oss-20b. Our platform was among the first to support both in production, so you can use them in existing workflows with or without OpenAI's API.
This post covers what the two models are, where you can run them, and how to switch an existing Bolder setup over.
The two models
gpt-oss-120b
- 120 billion parameters
- Built for advanced reasoning, planning and agentic applications
- Runs on data center hardware and high-end desktops
gpt-oss-20b
- 20 billion parameters
- Light enough for most desktops and laptops
What both models share
- Instruction following and chain-of-thought reasoning
- Native Python code execution
- Full-parameter fine-tuning for domain-specific tasks
- Safety training aligned with OpenAI's Preparedness Framework
- Permissive license for commercial use
- Available to download on Hugging Face and GitHub
Benchmarks
| Benchmark | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| MMLU | 90.0 | 85.3 |
| GPQA Diamond | 80.1 | 71.5 |
| AIME 2024 | 96.6 | 96.0 |
| AIME 2025 | 97.9 | 98.7 |
Running gpt-oss without an API
Open weights mean you can run OpenAI-level models on your own infrastructure instead of calling an API. Our deployment and integration layers work the same either way.
You can self-host gpt-oss on:
- Bolder-managed infrastructure, already tuned for these models
- Your own cloud or on-prem servers
- Edge and mobile environments (20B only)
What you get from self-hosting:
- Lower cost: no usage-based API billing
- Control: you decide latency, data residency and compliance
- Flexibility: you can fine-tune and modify the models for your domain
Switching models in existing products
If you use our copilots, RAG applications or Bolder Agents, you can swap these models in with little or no code change. The runtime supports:
- Model switching through config
- Centralized observability and logging
- Backward compatibility with both API-based and weights-based models
Bolder Agents on gpt-oss
Bolder Agents, our orchestration framework for AI automation, supports both gpt-oss models out of the box. It includes:
- Multi-step reasoning with constraint handling
- High-recall retrieval over enterprise datasets
- Domain-specific memory and context chaining
- Private, secure deployment with optional tool integrations
The models improve planning and execution accuracy on work like document analysis and customer support automation.
Getting started
Existing customers can use the models today:
- Pick a deployment mode:
- Keep using the OpenAI API
- Switch to self-hosted gpt-oss models
- Deploy gpt-oss inside your existing infrastructure
- Update your runtime configuration to enable the new models.
- Contact our solutions team if you want help with fine-tuning or infrastructure setup.
Documentation and reference deployments are in your dashboard.
Why open weights matter
Until now, using OpenAI's models meant depending on a proprietary API. With these weights, teams can run and fine-tune capable models directly, under a clear commercial license.
The gpt-oss models are fast, accurate, safety-evaluated and flexible enough for both research and production work. That covers copilots, document agents and internal LLM tooling.
If you are planning a move to self-hosted models, contact us and we will help you work through it.