Parameters
116.8B total / 5.1B active
MoE: total / active
Architecture
Mixture-of-Experts (MoE) Transformer
Released
04.08.2025
License
Apache License 2.0
Input Modalities
Output Modalities
Context (native)
131,072 tokens
Context (extended)
131,072 tokens
About
gpt-oss-120b (openai/gpt-oss-120b) is OpenAI's larger open-weight model, released August 4, 2025 under Apache 2.0 - a 116.83B-parameter Mixture-of-Experts transformer with 5.13B activated parameters per token, a 131,072-token context, and text-only input/output. The MoE weights are post-trained with MXFP4 quantization, so the model runs on a single 80GB GPU (NVIDIA H100 or AMD MI300X); all reported evals use that same MXFP4 quantization.
It offers configurable reasoning effort (low, medium, high) tuned to latency needs, exposes its full chain-of-thought (intended for debugging and trust, not for end users), and carries native agentic capabilities: function calling, web browsing, Python code execution and Structured Outputs. It was trained on a text-only dataset of trillions of tokens focused on STEM, coding and general knowledge (knowledge cutoff June 2024), with pre-training data filtered for harmful content using CBRN pre-training filters from GPT-4o. The models are fully fine-tunable. Model card paper: arXiv 2508.10925.
Training Data Trained on a text-only dataset with trillions of tokens, with a focus on STEM, coding, and general knowledge. Knowledge cutoff: June 2024. Pre-training data filtered for harmful content using CBRN pre-training filters from GPT-4o.
Benchmark Scores
| Benchmark | Score | Date |
|---|---|---|
|
AIME 2024
stem_reasoning
|
100.00%
|
08.08.2025 |
|
GPQA Diamond
stem_reasoning
|
73.15%
|
08.08.2025 |
|
Humanity's Last Exam
stem_reasoning
|
24.24%
|
08.08.2025 |
|
HLE (with tools)
stem_reasoning
|
3.60%
|
08.08.2025 |
|
MMLU
knowledge
|
94.41%
|
08.08.2025 |
|
SWE-bench Verified
coding_agent
|
71.10%
|
08.08.2025 |
|
Tau-Bench Retail
general_agent
|
87.84%
|
08.08.2025 |
|
Tau-Bench Airline
general_agent
|
100.00%
|
08.08.2025 |
|
Aider Polyglot
coding_agent
|
100.00%
|
08.08.2025 |
|
MMMLU
multilingual
|
60.70%
|
08.08.2025 |
|
HealthBench Hard
general_capabilities
|
100.00%
|
08.08.2025 |
|
HealthBench Consensus
general_capabilities
|
100.00%
|
08.08.2025 |
|
CodeForces
stem_reasoning
|
70.01%
|
08.08.2025 |
|
HealthBench
general_capabilities
|
85.31%
|
08.08.2025 |
|
AIME 2025
stem_reasoning
|
93.46%
|
08.08.2025 |
Architecture
- Attention
- Hybrid Attention (64:8)
- MoE
- 128 experts · top-4 per token
- Layers
- 36
- Hidden size
- 2880
- Context
- 131K tokens
- RoPE θ
- 150K
- Parameters
- 116830M
- Active params
- 5130M
Source: Hugging Face config.json · GptOssForCausalLM · exact layer pattern · model repo
Training Pipeline
-
1
pretraining
Pretraining
Trained on a text-only dataset with trillions of tokens, focusing on STEM, coding, and general knowledge. Knowledge cutoff: June 2024. Data filtered for harmful content using CBRN pre-training filters from GPT-4o. Training used NVIDIA H100 GPUs with PyTorch framework and expert-optimized Triton kernels. Required 2.1 million H100-hours. Uses Flash Attention algorithms. o200k_harmony tokenizer (BPE, 201,088 tokens).
-
2
sft
Post-Training: Distillation and SFT
Post-training uses similar CoT RL techniques as OpenAI o3. Models trained on harmony chat format with System > Developer > User > Assistant > Tool instruction hierarchy. Training dataset covers coding, math, science, and more.
-
3
rl
Reinforcement Learning for Reasoning and Tool Use
CoT RL techniques teach models to reason and solve problems using chain-of-thought. Variable effort reasoning training supports low, medium, high reasoning levels. Agentic tool use training includes browsing tool (search and open), Python tool (stateful Jupyter notebook), and arbitrary developer functions with structured outputs.
-
4
other
Safety Training: Deliberative Alignment
Deliberative alignment training teaches models to refuse disallowed content, be robust to jailbreaks, and adhere to instruction hierarchy. CBRN pre-training filters applied. Safety training follows OpenAI safety policies by default.
Training & Evaluation Datasets
| Name | Role | Size | Modalities | Collection |
|---|---|---|---|---|
| Text-only pre-training corpus (STEM, coding, general knowledge) | pretraining | — | — |
Linked Resources
gpt-oss-120b & gpt-oss-20b Model Card
https://arxiv.org/abs/2508.10925
OpenAI gpt-oss - Reference implementations and inference tools
https://github.com/openai/gpt-oss
OpenAI harmony - Harmony chat format implementation
https://github.com/openai/harmony
OpenAI TikToken - o200k_harmony tokenizer
https://github.com/openai/tiktoken
Introducing gpt-oss - OpenAI Blog
https://openai.com/index/introducing-gpt-oss/
Try gpt-oss
https://gpt-oss.com/
OpenAI Cookbook - gpt-oss Guides
https://cookbook.openai.com/topic/gpt-oss
gpt-oss Collection on HuggingFace
https://huggingface.co/collections/openai/gpt-oss
OpenAI Open Models
https://openai.com/open-models
awesome-gpt-oss - Collection of gpt-oss resources and inference partners
https://github.com/openai/gpt-oss/blob/main/awesome-gpt-oss.md
Trend Analysis
24h Change
+0.1%
7d Change
+1.0%
Current
40,435
downloads
+1.1%
likes
+0.1%
downloads
+0.8%
downloads_all_time
+0.3%
Usage & Social Metrics
| Source | Metric | Value | Period | Recorded |
|---|---|---|---|---|
| ollama | downloads | 12,500,000 pulls | daily | 01.09.2026 |
| huggingface | downloads_all_time | 53,799,428 | daily | 01.09.2026 |
| huggingface | followers | 40,435 | daily | 01.09.2026 |
| huggingface | likes | 5,141 | daily | 01.09.2026 |
| huggingface | downloads | 5,376,994 | daily | 01.09.2026 |
| ollama | downloads | 12,400,000 pulls | daily | 31.08.2026 |
| huggingface | downloads_all_time | 53,628,561 | daily | 31.08.2026 |
| huggingface | followers | 40,378 | daily | 31.08.2026 |
| huggingface | likes | 5,136 | daily | 31.08.2026 |
| huggingface | downloads | 5,316,895 | daily | 31.08.2026 |
| ollama | downloads | 12,400,000 pulls | daily | 30.08.2026 |
| huggingface | downloads_all_time | 53,510,645 | daily | 30.08.2026 |
| huggingface | followers | 40,323 | daily | 30.08.2026 |
| huggingface | likes | 5,136 | daily | 30.08.2026 |
| huggingface | downloads | 5,314,695 | daily | 30.08.2026 |
| ollama | downloads | 12,400,000 pulls | daily | 29.08.2026 |
| huggingface | followers | 40,265 | daily | 29.08.2026 |
| huggingface | likes | 5,133 | daily | 29.08.2026 |
| huggingface | downloads | 5,173,458 | daily | 29.08.2026 |
| ollama | downloads | 12,300,000 pulls | daily | 28.08.2026 |
| huggingface | followers | 40,216 | daily | 28.08.2026 |
| huggingface | likes | 5,130 | daily | 28.08.2026 |
| huggingface | downloads | 5,023,485 | daily | 28.08.2026 |
| ollama | downloads | 12,300,000 pulls | daily | 27.08.2026 |
| huggingface | followers | 40,157 | daily | 27.08.2026 |
| huggingface | likes | 5,128 | daily | 27.08.2026 |
| huggingface | downloads | 5,173,082 | daily | 27.08.2026 |
| ollama | downloads | 12,200,000 pulls | daily | 26.08.2026 |
| huggingface | followers | 40,076 | daily | 26.08.2026 |
| huggingface | likes | 5,126 | daily | 26.08.2026 |
| huggingface | downloads | 5,222,602 | daily | 26.08.2026 |
| ollama | downloads | 12,200,000 pulls | daily | 25.08.2026 |
| huggingface | followers | 40,023 | daily | 25.08.2026 |
| huggingface | likes | 5,125 | daily | 25.08.2026 |
| huggingface | downloads | 5,121,083 | daily | 25.08.2026 |
| ollama | downloads | 12,200,000 pulls | daily | 24.08.2026 |
| huggingface | followers | 39,962 | daily | 24.08.2026 |
| huggingface | likes | 5,123 | daily | 24.08.2026 |
| huggingface | downloads | 5,013,805 | daily | 24.08.2026 |