o3-mini

À propos de OpenAI

Fondée 2015 · San Francisco, USA

OpenAI was founded in 2015 as a non-profit lab by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba and John Schulman, transitioning to a capped-profit structure in 2019 to raise capital from Microsoft (now over $13B invested). OpenAI's research output includes the GPT series (GPT-1 in 2018 through GPT-4 in 2023), the InstructGPT/RLHF paper (2022), the GPT-4o System Card (2024) and the o-series reasoning papers. The o-series began with o1-preview in September 2024 and o1 in December 2024, introducing test-time compute scaling via long internal chains-of-thought trained with reinforcement learning. o3 and o3-mini followed in January-February 2025, with o3-mini specifically launched on 31 January 2025 as a fast, cost-efficient reasoning model available to free, Plus and Team ChatGPT users. The o-series has been described by OpenAI as the first models to solve previously-unsolved Frontier Math benchmark problems. Sam Altman remains CEO and the company's 2025 valuation exceeded $300 billion.

Architecture

Decoder-only Transformer (reasoning model with test-time chain-of-thought)

o3-mini is the small reasoning model in the o-series, released by OpenAI on 31 January 2025. Like its o1 and o3 siblings it is a decoder-only Transformer trained to produce long internal chains-of-thought during inference, then return a concise final answer to the user. Reinforcement learning was the core training technique: the model is rewarded for correct answers on verifiable tasks (math, code, scientific problems) and learns to allocate test-time compute by generating longer or shorter reasoning traces as needed. o3-mini supports three reasoning effort settings ('low', 'medium', 'high') that developers can choose to trade latency for accuracy. The training corpus and base pretraining are not disclosed in detail but build on OpenAI's frontier pretraining stack. o3-mini is the first o-series model to support function calling, Structured Outputs with strict JSON schemas, streaming and developer messages from day one, making it deployable in production agentic stacks. On benchmarks the medium-effort setting matched o1's performance at significantly lower cost; the high-effort setting outperformed o1 on AIME math, GPQA Diamond and Codeforces. Vision input is not supported (use o1 or GPT-4o for multimodal). The model retains OpenAI's deliberative alignment training, in which the safety policy itself is reasoned about explicitly during inference.

Paramètres

Undisclosed (estimated tens of billions of active parameters)

Contexte

200 000 tokens

Capacités

• Three reasoning effort modes: low, medium, high
• Strong math performance: o3-mini-high matches or beats o1 on AIME and GPQA
• Competitive Codeforces ratings on coding contests
• 200K context window with 100K max output
• Function calling and parallel tool calls
• Structured Outputs with strict JSON schemas
• Streaming and developer-message support
• Lower latency than o1 at equal accuracy
• Available in ChatGPT Free, Plus, Team and API
• Deliberative alignment: reasons about safety policy at inference time
• Best for: math, coding contests, scientific reasoning, multi-step tool agents.

Entraînement et licence

Pretrained on OpenAI's standard multi-trillion-token corpus (composition undisclosed). Post-training relies heavily on reinforcement learning against verifiable rewards on math, code and scientific problem datasets, plus deliberative alignment training where the model is taught to reason about OpenAI's safety policy.

Licence: Proprietary commercial license via OpenAI API, Azure OpenAI Service and ChatGPT. Commercial use permitted.

Tests de sécurité: Evaluated under OpenAI's Preparedness Framework. Deliberative alignment training reduces compliance with jailbreaks; external red-teamers covered persuasion, cyber and CBRN risks.