GPT OSS 120B
OpenAI's open-source 120B model with hybrid reasoning, extended thinking, efficient code generation, agentic search, computer use, and tool use capabilities.
GPT OSS 120B is OpenAI's first open-weight language model since GPT-2, released in August 2025 under the Apache 2.0 license. It is a 117B-parameter Mixture-of-Experts (MoE) transformer that activates only 5.1B parameters per forward pass, making it deployable on a single 80GB GPU (NVIDIA H100 or AMD MI300X). The model was post-trained using the same chain-of-thought reinforcement learning techniques as OpenAI's o-series reasoning models, enabling configurable reasoning effort (low / medium / high), full CoT visibility, tool use, and Structured Outputs.
In production, GPT OSS 120B is a strong contender for high-reasoning, agentic workloads where self-hosting is required — think on-premises enterprise deployments, sensitive-data processing, or government installations. It achieves near-parity with OpenAI o4-mini on core reasoning benchmarks while outperforming o3-mini across nearly every evaluation. The model is available via the OpenAI Responses API and through TheRouter's routing infrastructure for those who prefer a managed API experience while retaining the open-weight option.
- • High-reasoning self-hosted deployments — deploy on-premises on a single H100 or MI300X with full data sovereignty.
- • Agentic tool use workflows — function calling, web search integration, Python code execution within agentic loops.
- • Competitive programming and math — Codeforces Elo ~2620 and AIME 2024 with tools at 96.6%.
- • Medical and health reasoning — nearly matches o3 on HealthBench, outperforming o1, o4-mini, and GPT-4o.
- • Fine-tuning for specialized domains — Apache 2.0 license allows unfettered customization.
- • Latency-sensitive real-time applications — the high-reasoning mode produces long CoT traces; use gpt-oss-20b with low reasoning effort instead.
- • Multimodal tasks — this is a text-only model. For vision or audio, use GPT-4o or GPT-4o-mini via TheRouter.
- • Edge or mobile deployment — 60.8 GiB checkpoint is too large for most edge devices; use gpt-oss-20b (12.8 GiB) instead.
Modalities
Pricing Breakdown
| Type | Rate |
|---|---|
| Input | $0.162 / 1M tokens |
| Output | $0.648 / 1M tokens |
Supported Parameters
Specifications
| Release date | 2025-08-05openai.com ↗ | verified |
| Architecture | MoE Transformer — 36 layers, 128 experts, top-4 active, 116.8B total / 5.1B active paramsarxiv.org ↗ | verified |
| License | Apache 2.0openai.com ↗ | verified |
| Knowledge cutoff | June 2024arxiv.org ↗ | verified |
| Tokenizer | o200k_harmony (201,088 tokens)arxiv.org ↗ | verified |
| Quantization | MXFP4 (4.25 bits/param on MoE weights)arxiv.org ↗ | verified |
| Checkpoint size | 60.8 GiBarxiv.org ↗ | verified |
| Training compute | 2.1 million H100-hoursarxiv.org ↗ | verified |
| Reasoning modes | low / medium / high (configurable in system prompt)openai.com ↗ | verified |
Benchmarks
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MMLU High reasoning mode | 90.0%% | arxiv.org ↗ | |
AIME 2024 With tools, high reasoning mode | 96.6%% | arxiv.org ↗ | |
AIME 2025 With tools, high reasoning mode | 97.9%% | arxiv.org ↗ | |
Codeforces Elo Estimated Elo rating; high reasoning mode with terminal tool | ~2620 | openai.com ↗ | |
SWE-Bench Verified High reasoning mode | 62.4%% | arxiv.org ↗ | |
GPQA Diamond With tools, high reasoning mode | 80.9%% | arxiv.org ↗ | |
HLE With tools, high reasoning mode. HLE is an extremely difficult benchmark. | 19.0%% | arxiv.org ↗ | |
HealthBench High reasoning mode. Nearly matches o3, outperforms o1, GPT-4o, o3-mini, and o4-mini. | 57.6%% | arxiv.org ↗ | |
HealthBench Hard High reasoning mode, challenging subset of HealthBench | 30.0%% | arxiv.org ↗ | |
Tau-Bench Retail High reasoning mode — agentic function calling evaluation | 67.8%% | arxiv.org ↗ | |
MMMLU (Average) 14-language average, high reasoning mode | 81.3%% | arxiv.org ↗ |
API Usage Examples
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "openai/gpt-oss-120b",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'API guide
Full API reference →Chat completion
Use standard OpenAI-compatible chat completion via TheRouter. The model uses the harmony chat format; TheRouter handles the conversion automatically.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-oss-120b",
"messages": [
{"role": "user", "content": "Explain the benefits of Mixture-of-Experts architecture in large language models."}
]
}'More from openai
Similar models
Cross-provider sibling modelsFrequently asked
Can I use GPT OSS 120B through TheRouter's API without self-hosting?
Yes. TheRouter routes requests to gpt-oss-120b through its managed infrastructure. You get the open-weight model via a standard OpenAI-compatible API at https://api.therouter.ai/v1, without managing your own GPU hardware.
What license does GPT OSS 120B use?
Apache 2.0 — a permissive license that allows commercial use, modification, distribution, and sublicensing without copyleft restrictions or patent risk.
How does GPT OSS 120B compare to GPT-4o?
GPT OSS 120B is a reasoning-focused open-weight model that outperforms GPT-4o on math (AIME), coding (Codeforces), and health (HealthBench) benchmarks. However, GPT-4o is a multimodal model (vision, audio) while GPT OSS 120B is text-only. For multimodal tasks, use GPT-4o or GPT-4o-mini via TheRouter.
What hardware do I need to self-host GPT OSS 120B?
A single 80GB GPU such as NVIDIA H100 or AMD MI300X. The quantized checkpoint is 60.8 GiB. With MXFP4 quantization and MoE architecture, only 5.1B parameters are active per forward pass, enabling efficient inference on consumer GPU hardware.
Fact ledger — every claim on this page traces here
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com ↗ | 2026-05-29 | verified |
| Architecture | arxiv.org ↗ | 2026-05-29 | verified |
| License | openai.com ↗ | 2026-05-29 | verified |
| Knowledge cutoff | arxiv.org ↗ | 2026-05-29 | verified |
| Tokenizer | arxiv.org ↗ | 2026-05-29 | verified |
| Quantization | arxiv.org ↗ | 2026-05-29 | verified |
| Checkpoint size | arxiv.org ↗ | 2026-05-29 | verified |
| Training compute | arxiv.org ↗ | 2026-05-29 | verified |
| Reasoning modes | openai.com ↗ | 2026-05-29 | verified |
| MMLU | arxiv.org ↗ | 2026-05-29 | verified |
| AIME 2024 | arxiv.org ↗ | 2026-05-29 | verified |
| AIME 2025 | arxiv.org ↗ | 2026-05-29 | verified |
| Codeforces Elo | openai.com ↗ | 2026-05-29 | to verify |
| SWE-Bench Verified | arxiv.org ↗ | 2026-05-29 | verified |
| GPQA Diamond | arxiv.org ↗ | 2026-05-29 | verified |
| HLE | arxiv.org ↗ | 2026-05-29 | verified |
| HealthBench | arxiv.org ↗ | 2026-05-29 | verified |
| HealthBench Hard | arxiv.org ↗ | 2026-05-29 | verified |
| Tau-Bench Retail | arxiv.org ↗ | 2026-05-29 | verified |
| MMMLU (Average) | arxiv.org ↗ | 2026-05-29 | verified |
| Can I use GPT OSS 120B through TheRouter's API without self-hosting? | openai.com ↗ | 2026-05-29 | to verify |
| What license does GPT OSS 120B use? | openai.com ↗ | 2026-05-29 | to verify |
| How does GPT OSS 120B compare to GPT-4o? | arxiv.org ↗ | 2026-05-29 | to verify |
| What hardware do I need to self-host GPT OSS 120B? | github.com ↗ | 2026-05-29 | to verify |