GPT-OSS-20B — Model Documentation
Published by Nextbit256 S.L. | Last updated: May 2026 Compliance reference: AI Act Regulation (EU) 2024/1689, Art. 53.1.c and 53.1.d
Model Identification
| Field | Detail |
|---|---|
| Full model name | GPT-OSS-20B |
| Developed by | OpenAI |
| Release date | August 2025 |
| Model card | huggingface.co/openai/gpt-oss-20b |
| Technical report | arXiv:2508.10925 — gpt-oss-120b & gpt-oss-20b Model Card |
| Developer blog | Not published by OpenAI for this model |
Architecture and Parameters
| Field | Detail |
|---|---|
| Architecture | Mixture of Experts (MoE) transformer |
| Total parameters | ~21B (published as "21B" and "22B" in different sources) |
| Active parameters per token | 3.6B |
| Non-embedding parameters | Not published by OpenAI |
| Layers | Not published by OpenAI |
| Attention | Not published by OpenAI |
| Native context length | Not published by OpenAI |
| Extended context | Not published by OpenAI |
| Precision | BF16 (full precision); U8 and MXFP4 (MoE weight quantization available) |
Note: The model uses MXFP4 quantization of MoE weights, enabling it to run within 16 GB of memory despite the total parameter count.
License and Commercial Use
License: Apache License 2.0
Apache 2.0 is a permissive open-source license that allows free commercial use, distribution, and modification, provided that the original copyright notice and license text are retained and any modifications are documented.
Nextbit's verification: Nextbit has reviewed the license terms applicable to GPT-OSS-20B and confirmed that serving this model via API under a commercial inference service is permitted under Apache 2.0. No separate commercial agreement with OpenAI is required for this use case.
Restrictions relevant to users: None specific to this model beyond standard Apache 2.0 terms. The model was trained on OpenAI's internal harmony response format, but no IP or license restrictions arise from this for API users. Users of Nextbit's API who integrate model outputs into their own products remain responsible for their own compliance with applicable law and any downstream licensing requirements.
Training Data Summary
The following is based on publicly available information from the model card (arXiv:2508.10925) and the HuggingFace model card.
Pretraining and distillation data
GPT-OSS-20B was trained using large-scale distillation and reinforcement learning. OpenAI describes the model as an open-weight reasoning model trained with distillation from larger internal models and RL-based fine-tuning.
The model was trained using OpenAI's internal harmony response format.
Token volume: Not published by OpenAI.
Languages: Not explicitly published by OpenAI. Given the model's position as a general-purpose reasoning model, multilingual support is expected but not formally documented.
Knowledge cutoff: Not published by OpenAI.
What is not publicly available
OpenAI has not published:
- Training data sources or corpora
- Total pretraining token count
- Language distribution
- Knowledge cutoff date
- Training compute (FLOPs or GPU-hours)
- Detailed architecture specifications (layers, attention heads, expert routing)
The primary public reference is the technical report arXiv:2508.10925 and the HuggingFace model card. Nextbit will update this section if OpenAI publishes additional training data documentation.
Languages Supported
Not explicitly published by OpenAI. Refer to the HuggingFace model card and the technical report for any updates.
Intended Uses
This model is designed for:
- Reasoning-intensive tasks with configurable effort levels (Low, Medium, High reasoning effort settings)
- Full chain-of-thought access — the complete reasoning process is visible, not hidden
- Fine-tuning on consumer hardware (the model is explicitly designed to be fine-tunable at low resource cost)
- Agentic capabilities: function calling, web browsing, Python code execution, and Structured Outputs
- Applications where an open-weight, fine-tunable reasoning model is preferred over closed API access
No explicitly excluded uses are stated in the model card beyond standard responsible use guidelines.
Systemic Risk Assessment
| Field | Detail |
|---|---|
| Training FLOPs | Not published by OpenAI |
| Estimated FLOPs (active params method) | ~1.9 × 10²³ (estimate based on 3.6B active parameters × an assumed ~9T training tokens; formula: 6 × N_active × D; both active parameter count and token count are estimates from public sources) |
| Exceeds 10²⁵ FLOPs threshold? | No — estimated to be well below the systemic risk threshold |
| Note on MoE architecture | FLOPs for MoE models are calculated based on active parameters per token, not total parameters. The 3.6B active parameter figure (not the ~21B total) is the relevant input for FLOPs estimation. |
| AI Office designation | Not designated as a systemic risk model as of May 2026 |
| Art. 55 obligations apply? | No |
This estimate is highly uncertain because OpenAI has not published training token counts, architecture details, or compute figures for GPT-OSS-20B. The active parameter count of 3.6B is taken from the model card; the token count is an assumption. If OpenAI publishes official figures, this section will be updated.
This documentation is published by Nextbit256 S.L. in accordance with Article 53(1)(c) and 53(1)(d) of Regulation (EU) 2024/1689 (AI Act). Nextbit256 S.L. serves this model via its inference API but did not develop or train it. All information is sourced from public documentation published by OpenAI.
For questions: [email protected]