The new R1-0528 version lifts math performance from 70% to 87.5% on the AIME 2025 benchmark and cuts hallucinations nearly in half, bringing DeepSeek's open model closer to o3 and Gemini 2.5 Pro.
Anthropic unveils Claude 3.7 Sonnet, which combines instant answers with visible step-by-step reasoning in a single model, alongside Claude Code, a coding agent that works from the terminal.
Alibaba unveils Qwen2.5-Max, an MoE model trained on more than 20 trillion tokens that the company says outperforms DeepSeek-V3 and GPT-4o on several benchmarks. The announcement lands squarely in the middle of the DeepSeek fallout and the Chinese New Year.
Chinese lab DeepSeek has released R1, a reasoning model matching OpenAI's o1, with open weights under an MIT license, six distilled models and a far cheaper API.
Microsoft has unveiled Phi-4, a 14-billion-parameter model that prioritizes mathematical reasoning over sheer size, trained largely on synthetic data designed to maximize learning quality.
Alibaba's Qwen team has released QwQ-32B-Preview, a 32-billion-parameter model that reasons step by step before answering. It's open-access and arrives just months after OpenAI unveiled o1.
DeepSeek has unveiled R1-Lite-Preview, a model designed to spend more steps on math and logic problems. The Chinese company says it comes close to OpenAI’s o1-preview reasoning model.
Anthropic updates Claude 3.5 Sonnet, introduces Claude 3.5 Haiku and makes available in beta a tool that lets the model use graphical interfaces with a cursor, clicks and a keyboard.
Alibaba has released Qwen 2.5, a family of open language models ranging from 500 million to 72 billion parameters. The release expands the options for running generative AI at very different costs and capability levels.
OpenAI introduces o1-preview and o1-mini, two models designed to spend more time on internal reasoning. The company is targeting math, science, and programming, although the launch comes with usage limits and missing features.
Anthropic is making prompt caching for Claude available in public beta. The feature reuses large parts of a request, cuts input costs by up to 90% and avoids processing the same context on every call.
OpenAI introduces GPT-4o mini, a lightweight model for high-volume tasks that improves on GPT-3.5 Turbo while cutting usage costs. It comes to ChatGPT and the API with text and vision capabilities.
This website uses cookies to improve the browsing experience. Cookie policy.
