gpt-oss (OpenAI open-weight models)

gpt-oss-120b and gpt-oss-20b are OpenAI's first open-weight reasoning models, published on August 5, 2025 with full weights released on Hugging Face under the Apache 2.0 license (Source: https://openai.com/index/introducing-gpt-oss/). They were OpenAI's first open-weight language-model release since GPT-2 in 2019. Altman positioned gpt-oss-120b as a "state-of-the-art open-weights reasoning model, with strong real-world performance comparable to o4-mini" (Altman, Aug 5, 2025); independent testers subsequently disputed how far the benchmark results carried over to real-world use (see Reception). The full model card is summarized at OpenAI gpt-oss Model Card (August 2025). In October 2025 OpenAI extended the family with gpt-oss-safeguard, a pair of fine-tuned safety-classification models (Source: https://openai.com/index/introducing-gpt-oss-safeguard/).

Lineage and history

Although OpenAI had released other models openly, including Whisper and CLIP, the gpt-oss models were its first open-weight language models since GPT-2 (2019) (Source: https://openai.com/index/introducing-gpt-oss/). Altman first announced an intent to release an open-weight reasoning model on March 31, 2025. On June 10, 2025, the release was pushed to "later this summer but not June", with Altman saying the "research team did something unexpected and quite amazing." On July 11, 2025, Altman delayed the release further, citing additional safety tests and review of high-risk areas. Both gpt-oss-120b and gpt-oss-20b were released on August 5, 2025.

The release placed OpenAI within the open-weight segment alongside Meta's Llama family, DeepSeek, Moonshot's Kimi K2, Alibaba's Qwen, and Mistral, and reduced the framing of OpenAI as an exclusively closed-model developer. OpenAI positioned gpt-oss as comparable to o4-mini rather than to the GPT-5 family, so the release does not indicate OpenAI's frontier model direction. Through gpt-oss-20b's claimed ability to run locally on a phone (Altman described gpt-oss-20b as runnable on a phone), the release carries distribution implications for on-device deployment.

On October 29, 2025, OpenAI released a research preview of gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, fine-tuned versions of the gpt-oss models for safety-classification tasks, under the same Apache 2.0 license (Source: https://openai.com/index/introducing-gpt-oss-safeguard/).

Architecture and training

Both models are autoregressive Transformers using a mixture-of-experts (MoE) design to reduce active parameters per token. gpt-oss-120b has 116.8B total parameters (5.13B active per token) with 128 experts and top-4 routing across 36 layers; gpt-oss-20b has 20.9B total parameters (3.61B active) with 32 experts and top-4 routing across 24 layers (Source: https://openai.com/index/introducing-gpt-oss/; gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, August 2025)). The models use alternating dense and locally banded sparse attention patterns similar to GPT-3, grouped multi-query attention with a group size of 8, Rotary Positional Embedding (RoPE), and a native context length of 128k tokens (Source: https://openai.com/index/introducing-gpt-oss/).

Both models apply MXFP4 quantization to the MoE weights at 4.25 bits per parameter. With this quantization, gpt-oss-120b fits on a single 80GB GPU, and gpt-oss-20b runs on systems with as little as 16GB of memory.

Pre-training used a mostly English, text-only dataset focused on STEM, coding, and general knowledge, with certain Chemical, Biological, Radiological, and Nuclear (CBRN)-related data filtered out. The data was tokenized with o200k_harmony, a superset of the tokenizer used for o4-mini and GPT-4o, which OpenAI open-sourced alongside the models (Source: https://openai.com/index/introducing-gpt-oss/). Post-training followed a process similar to o4-mini's — a supervised fine-tuning stage and a high-compute reinforcement-learning stage — aligned to the OpenAI Model Spec. The models support three reasoning-effort settings (low, medium, high) set via the system message, and were post-trained on OpenAI's "harmony" prompt format; OpenAI open-sourced a harmony renderer in Python and Rust, plus reference inference implementations for PyTorch and Apple Metal (Source: https://openai.com/index/introducing-gpt-oss/).

OpenAI applied no direct supervision to either model's chain-of-thought, in line with its stated position that unsupervised CoT is needed to monitor model misbehavior, deception, and misuse; it cautions that developers should not display raw CoT to end users, since it may contain hallucinated or harmful content (Source: https://openai.com/index/introducing-gpt-oss/).

Capabilities and benchmarks

OpenAI-reported evaluations at release (August 5, 2025) compared the models to its proprietary reasoning models (Source: https://openai.com/index/introducing-gpt-oss/):

\* o3-mini was evaluated without tools. Without tools, gpt-oss-120b scored 2463 Codeforces Elo and 14.9% on Humanity's Last Exam; gpt-oss-20b scored 2230 and 10.9%.

On this basis OpenAI stated that gpt-oss-120b achieves near-parity with o4-mini on core reasoning benchmarks and outperforms it on HealthBench and competition mathematics, while gpt-oss-20b matches or exceeds o3-mini despite running in 16GB of memory (Source: https://openai.com/index/introducing-gpt-oss/). Both models are designed for agentic workflows, with instruction following, tool use (web search, Python code execution), full chain-of-thought, and Structured Outputs support.

An independent academic evaluation (Bi et al., arXiv:2508.12461, August 2025) tested both models unquantized against six contemporary open-weight models from 14.7B to 235B parameters across ten benchmarks and found both "mid-tier" within the open-source landscape, with relative strength in code generation and notable weakness in multilingual tasks; it also found that gpt-oss-20b consistently outperformed gpt-oss-120b on several benchmarks, including HumanEval and MMLU, despite requiring substantially less memory (Source: https://arxiv.org/abs/2508.12461).

Availability and licensing

The models are released under Apache 2.0 plus a gpt-oss usage policy, which OpenAI characterized as more permissive than DeepSeek's custom license. Weights for gpt-oss-120b are at https://huggingface.co/openai/gpt-oss-120b, natively quantized in MXFP4.

OpenAI partnered ahead of launch with deployment platforms including Azure, Hugging Face, vLLM, Ollama, llama.cpp, LM Studio, AWS, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, and OpenRouter, and with hardware makers NVIDIA, AMD, Cerebras, and Groq (Source: https://openai.com/index/introducing-gpt-oss/). Both models became available on Amazon Bedrock and SageMaker JumpStart on release day (Source: https://aws.amazon.com/blogs/aws/openai-open-weight-models-now-available-on-aws/), and Cerebras offered day-one hosted inference, stating that gpt-oss-120b ran fastest on its hardware (Source: https://www.cerebras.ai/blog/openai-gpt-oss-120b-runs-fastest-on-cerebras). Microsoft shipped GPU-optimized ONNX Runtime versions of gpt-oss-20b for Windows via Foundry Local and the AI Toolkit for VS Code. OpenAI hosts a browser playground at gpt-oss.com and said it might consider API support for gpt-oss in the future; early enterprise partners included AI Sweden, Orange, and Snowflake (Source: https://openai.com/index/introducing-gpt-oss/).

Safety and evaluations

The models received safety training including CBRN data filtering during pre-training and, during post-training, deliberative alignment and the instruction hierarchy to teach refusal of unsafe prompts and defense against prompt injection (Source: https://openai.com/index/introducing-gpt-oss/).

Because a released open-weight model can be fine-tuned by adversaries, OpenAI supplemented standard evaluations with a worst-case (malicious) fine-tuning assessment: it fine-tuned gpt-oss-120b on specialized biology and cybersecurity data — using its own training stack, the way an attacker might — to create a domain-specific non-refusing version for each domain, then evaluated those models internally and externally. OpenAI's Safety Advisory Group concluded that the model does not reach High capability in either Biological/Chemical Risk or Cyber risk even under adversarial fine-tuning (Source: gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, August 2025)). The methodology is detailed in an accompanying safety paper, "Estimating Worst-Case Frontier Risks of Open-Weight LLMs" (Source: https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms/), and was reviewed by three independent expert groups, many of whose recommendations OpenAI says it adopted (Source: https://openai.com/index/introducing-gpt-oss/).

Provenance note: the safety paper is queued for ingest as a foundational source (Wiki/_meta/queue/INGEST-estimating-worst-case-frontier-risks-open-weight-llms-2026-07-02.md); its conclusions are currently summarized in gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, August 2025).

OpenAI also concluded that the release of gpt-oss-120b does not significantly advance the frontier of biological capabilities in open foundation models, on the basis that existing open-weight models already approach the adversarially-fine-tuned performance of gpt-oss-120b on most evaluations. These findings informed the release decision. The safety case rests on the gpt-oss model card (oai_gpt-oss_model_card.pdf) and the OpenAI Preparedness Framework V.2.

Alongside the release, OpenAI hosted a red-teaming challenge on Kaggle with a $500,000 prize fund judged by experts from OpenAI and other labs, committing to publish a report and open-source an evaluation dataset from validated findings (Source: https://openai.com/index/introducing-gpt-oss/); winners were later announced on Kaggle (Source: https://www.kaggle.com/competitions/openai-gpt-oss-20b-red-teaming). Commentators at Tech Policy Press situated the challenge within the evolution of AI safety practices and the political and regulatory dynamics around open models, including the Trump administration's AI Action Plan (Source: https://techpolicy.press/what-openais-latest-redteaming-challenge-reveals-about-the-evolution-of-ai-safety-practices).

gpt-oss-safeguard

On October 29, 2025, OpenAI released a research preview of gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, open-weight reasoning models for safety-classification tasks fine-tuned from the corresponding gpt-oss models and released under the same Apache 2.0 license and gpt-oss usage policy (Source: https://openai.com/index/introducing-gpt-oss-safeguard/). Rather than training a classifier on labeled examples of a fixed policy, the models take a developer-provided policy and the content to classify as inputs at inference time, output a classification with chain-of-thought reasoning, and allow policies to be revised without retraining. The approach is an open-weight implementation of an internal OpenAI tool called Safety Reasoner, which OpenAI says has consumed as much as 16% of total compute in some recent launches and forms part of its safeguards for systems including image generation and Sora 2 (Source: https://openai.com/index/introducing-gpt-oss-safeguard/).

OpenAI developed the release with the safety-tooling organization ROOST, which launched a ROOST Model Community for open safety models alongside it; early testers included SafetyKit, Tomoro, and Discord. In OpenAI's reported evaluations, the safeguard models outperformed gpt-5-thinking and the base gpt-oss models on internal multi-policy accuracy (safeguard-120b 46.3%, safeguard-20b 43.6%, vs. internal Safety Reasoner 52.2%, gpt-5-thinking 43.2%, gpt-oss-120b 32.5%, gpt-oss-20b 32.1%) and slightly outperformed all tested models on the 2022 OpenAI moderation set (F1 82.9), while trailing Safety Reasoner and gpt-5-thinking marginally on ToxicChat (Source: https://openai.com/index/introducing-gpt-oss-safeguard/). OpenAI's accompanying technical report states two limitations — dedicated classifiers trained on tens of thousands of labeled samples can still outperform policy-reasoning classification, and the approach is compute-intensive — and notes that because the safeguard models were trained without additional biological or cybersecurity data, the gpt-oss worst-case fine-tuning analysis cross-applies (Source: https://openai.com/index/gpt-oss-safeguard-technical-report/).

Provenance note: the gpt-oss-safeguard technical report (PDF on cdn.openai.com, October 29, 2025) is queued for ingest as a foundational source (Wiki/_meta/queue/INGEST-gpt-oss-safeguard-technical-report-2026-07-02.md).

Ecosystem and adoption

OpenAI's Romain Huet reported over 5 million Hugging Face downloads and more than 400 fine-tuned derivatives within the first week, describing gpt-oss-120b as the fastest model to gain popularity on the platform (Source: https://www.linkedin.com/posts/romainhuet_over-5m-downloads-in-under-a-week-for-our-activity-7360755653113507843-fnq3). NVIDIA published quantization-aware-training guidance for fine-tuning gpt-oss while preserving MXFP4 efficiency (Source: https://developer.nvidia.com/blog/fine-tuning-gpt-oss-for-accuracy-and-performance-with-quantization-aware-training/).

Hugging Face's Spring 2026 ecosystem review grouped gpt-oss with AI2's OLMo and Google's Gemma as the main Western efforts to offer "commercially deployable alternatives to Chinese models", while reporting that Chinese models had reached 41% of platform downloads over the preceding year and that China had surpassed the United States in monthly and overall downloads; it called whether these Western efforts "can match the adoption momentum of Qwen and DeepSeek" a defining question of 2026 (Source: https://huggingface.co/blog/huggingface/state-of-os-hf-spring-2026).

Reception

Initial reviews from developers running the models locally were favorable. Simon Willison called the models "really good" and highlighted the Apache 2.0 licensing and small memory footprints (Source: https://simonwillison.net/2025/Aug/5/gpt-oss/). Nathan Lambert of Interconnects gave OpenAI "a very strong grade" for the release, called it "a phenomenal step for the open ecosystem, especially for the West and its allies", noted that the Apache 2.0 license leaves the models free for distillation and commercial deployment, and estimated that serving gpt-oss would cost at least ten times less than o4-mini's API pricing of $1.10 per million input tokens and $4.40 per million output tokens (Source: https://www.interconnects.ai/p/gpt-oss-openai-validates-the-open).

Subsequent independent testing was more critical, centering on a gap between benchmark scores and real-world performance. Zvi Mowshowitz's roundup of early evaluations reported that the models performed well in their targeted reasoning domains but struggled in many practical uses, collecting assessments that the models were "very, very benchmaxxed" (with third-party and private benchmarks placing gpt-oss-120b below o4-mini, below the largest Qwen releases, and at times below newer ~30B Qwen models), that high hallucination rates and thin world knowledge limited tool use — one commentator said the hallucination rates suggested gpt-oss was "close to Sam's vision of a platonic ideal of a 'very tiny reasoning model with no knowledge'" — and comparisons to Microsoft's Phi series ("Great benchmark scores, trained on lots of synthetic data, great at STEM, sucks at everything else") (Source: https://www.lesswrong.com/posts/AJ94X73M6KgAZFJH2/openai-s-gpt-oss-is-already-old-news). The researcher nostalgebraist described the models as "almost Llama 4 levels of bad", reporting brittleness outside the harmony chat format and that the model is "ignorant about its own ignorance" (Source: https://www.lesswrong.com/posts/AJ94X73M6KgAZFJH2/openai-s-gpt-oss-is-already-old-news). Others noted a confound: researcher Havard Ihle observed that some mediocre independent results may reflect misconfigured third-party inference providers rather than the models themselves (Source: https://www.lesswrong.com/posts/AJ94X73M6KgAZFJH2/openai-s-gpt-oss-is-already-old-news). The Bi et al. academic evaluation likewise placed both models mid-tier among open-weight peers (Source: https://arxiv.org/abs/2508.12461).

Role in open-weights policy debates

OpenAI framed the release in policy terms, writing that "broad access to these capable open-weights models created in the US helps expand democratic AI rails" and that open models lower barriers for emerging markets, resource-constrained sectors, and smaller organizations (Source: https://openai.com/index/introducing-gpt-oss/). The release followed the July 2025 White House America's AI Action Plan, which elevated open-weight models as a strategic asset for U.S. competitiveness (Source: https://hai.stanford.edu/assets/files/hai-digichina-issue-brief-beyond-deepseek-chinas-diverse-open-weight-ai-ecosystem-policy-implications.pdf). Lambert framed the release against the growth of the Chinese open-weight ecosystem (DeepSeek, Qwen, Kimi K2), writing that it "could be the start of the turning point of adoption and impact of open models relative to China" (Source: https://www.interconnects.ai/p/gpt-oss-openai-validates-the-open).

For the Open-Source AI / Open-Weight Models debate, the worst-case fine-tuning assessment supplied a documented methodology for arguing that a capable open-weight model can be released with accompanying evaluation evidence rather than withheld because downstream use cannot be controlled; the model card's frontier-advancement comparison against existing open models is the core of that argument (Source: gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, August 2025)). The release is also a central case for Open-Weight Frontier Models — a proprietary frontier lab publishing weights under a permissive license while reserving its frontier systems for API access.

The safeguard variants of the family have their own technical report (Technical Report: gpt-oss-safeguard-120b and gpt-oss-safeguard-20b (OpenAI, October 2025)), covering safety-classification performance and its stated limits, multilingual performance, and observed failures across disallowed content, jailbreaks, and instruction-hierarchy conflicts.

Relationships

• developer: OpenAI
• supports: Open-Weight Frontier Models, Open-Source AI / Open-Weight Models
• depends-on: gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, August 2025), OpenAI Preparedness Framework V.2
• related: OpenAI o-series (o1 → o4-mini), Llama 4 (Meta), DeepSeek-R1, Qwen3, Kimi K2, Gemma (Google open-weight models), America's AI Action Plan, Sam Altman.

Sources

• OpenAI / Introducing gpt-oss (2025-08-05): https://openai.com/index/introducing-gpt-oss/
• System Card: oai_gpt-oss_model_card.pdf (gpt-oss-120b & gpt-oss-20b Model Card (OpenAI, August 2025); arXiv:2508.10925)
• OpenAI / Estimating Worst-Case Frontier Risks of Open-Weight LLMs (2025-08-05): https://openai.com/index/estimating-worst-case-frontier-risks-of-open-weight-llms/
• OpenAI / Introducing gpt-oss-safeguard (2025-10-29): https://openai.com/index/introducing-gpt-oss-safeguard/
• OpenAI / gpt-oss-safeguard technical report (2025-10-29): https://openai.com/index/gpt-oss-safeguard-technical-report/
• Altman delay announcements (2025-06-10, 2025-07-11)
• Altman launch thread (2025-08-05)
• Bi et al., "Is GPT-OSS Good?" (arXiv:2508.12461)
• Hugging Face, "State of Open Source on Hugging Face: Spring 2026" (2026-03-17)
• Mowshowitz, "OpenAI's GPT-OSS Is Already Old News" (2025-08); Willison (2025-08-05); Lambert/Interconnects (2025-08-05)