Énorme victoire pour l’open source ! La nouvelle version de DeepSeek R1 rivalise désormais avec le modèle o3-high d’OpenAI sur LiveCodeBench.
OPEN SOURCE
-
Darwin Godel Machine: Self-Improving Agents Evolution
By
–
Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
Paper: https://
arxiv.org/pdf/2505.22954
.pdf
…
Code: https://
github.com/jennyzzt/dgm -
Fast-dLLM: Training-free Acceleration of Diffusion Language Models
By
–
Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding
Paper: https://
arxiv.org/pdf/2505.22618
.pdf
…
Code: https://
github.com/NVlabs/Fast-dL
LM
…
Project Page: https://
nvlabs.github.io/Fast-dLLM -
Mistral AI lance l’API Agents avec planification de tâches
By
–
ICYMI: Mistral AI released Agents API for public use with task planning and tool integration. Mistral Agents are now also compatible with MCPs!
— 🚨 AI News | TestingCatalog (@testingcatalog) 29 mai 2025
An agent makes PR to github demo 👀 https://t.co/R0jilWFB4S pic.twitter.com/Vy5HLu905bICYMI: Mistral AI released Agents API for public use with task planning and tool integration. Mistral Agents are now also compatible with MCPs! An agent makes PR to github demo
-
FLUX.1 Kontext outperforms OpenAI’s GPT-4o in image generation tests
By
–
In our tests we’ve found FLUX.1 Kontext to give accurate and brilliant results. It’s better and cheaper than OpenAI’s 4o/gpt-image-1 model (and there’s no yellow tint). We've put our findings into a blog post: https://
replicate.com/blog/flux-kont
ext
… An open-source dev version of Kontext is -
Distilled Qwen3-8B Model for Personal Computers
By
–
Como ocurrió con la versión anterior, el modelo original no está pensado para existir en nuestros modestos ordenadores (es muy grande). Por eso tenéis que optar por las versiones destiladas que en esta ocasión únicamente tenemos sobre Qwen3-8B Esa es la que hay que descargar
-
Work Modes Impact on Team Capacity and Productivity
By
–
This post by @johncutlefish differentiates work "modes" and how they affect capacity: there's trailblazing, spellcasting, questing, and mapchecking, to name a few! https://
bit.ly/43cosoQ #digitaltransformation #opensource #foss -
DeepSeek R1 Update Rivals Leading AI Models Performance
By
–
¡NUEVO MODELO DEEPSEEK R1! La ballena azul que nos hizo soñar con IAs opensource competitivas, acaba de actualizarse con un modelo aún más potente! Si miramos en diferentes benchmarks de referencia estamos ante un modelo que se queda muy cerca de los líderes o3 y Gemini Pro!
-
DeepSeek-R1-0528 evals released
By
–
DeepSeek-R1-0528 evals are out. Also confirmed that this upgrade powers DeepSeek chat and APIs.
-
DeepSeek R1 Gains Self-Correction and Creative Worldbuilding Capabilities
By
–
Le nouveau DeepSeek r1 est en fait plutôt bon. Il est désormais capable de corriger sa propre chaîne de pensée comme o3, et de faire du worldbuilding créatif comme Claude.
R1 n’en était pas capable auparavant. Ils vont bientôt publier les poids du modèle, pour que le monde