Disruption-compensation reset
The product lead apologized for a disruption and confirmed that everyone's usage limits had been reset.
SEARCH
Search English reports
English reports are listed here. AI-generated drafts are clearly labeled until an editor reviews them.
Imported · reviewedImpact 49/100
Xiaomi releases MiMo V2 Flash
Xiaomi open-sourced an efficient MoE base for code and agents under an MIT license, using hybrid attention and a multi-layer MTP design.
Allen Institute for AI releases Molmo 2
The second generation of the Allen Institute's lightweight, natively multimodal vision model.
OpenAI releases GPT-5.2
A follow-up update in the GPT-5 series that improves text diversity and further lowers the total cost of high-volume usage.
Mistral releases 24B coding model Devstral Small 2
Devstral Small 2 offers a lighter alternative to Devstral 2, sized for individual developers running coding assistance on consumer hardware.
Mistral AI releases Devstral 2
A specialized model from Mistral AI tuned for large legacy-codebase refactoring, long-context analysis, and low-level debugging.
Codex uses an all-user reset to mitigate an incident
To mitigate an incident quickly, the team resets limits for all Codex users and says distribution will complete within ten minutes.
Mistral AI releases Mistral Large 3
Mistral's third-generation large model, lowering compute cost by tuning per-forward active parameters while holding overall quality.
Amazon releases Nova 2
Amazon's second-generation model on a rebuilt multimodal architecture, with a substantial overall capability upgrade.
Mistral releases Ministral 3 models for edge deployment
Mistral releases the Ministral 3 family in 3B, 8B and 14B sizes for edge devices and compute-constrained environments.
DeepSeek releases DeepSeek V3.2
Another fine-tuning iteration on the V3 base, focused on lowering peak memory use and speeding up inference on local hardware.
Anthropic releases Claude Opus 4.5
A half-step 4.5 upgrade to Opus that strengthens deep reasoning while maintaining long-form generation quality.
OpenAI resets all Codex users’ limits after high latency near the United States
After several hours of unusually high latency in the United States and nearby regions, the team resets limits for all Codex users.
Ai2 releases fully open 32B foundation model OLMo 3
Ai2 releases OLMo 3, a 32-billion-parameter foundation model that continues its emphasis on open training data and architecture details.
Deep Cogito releases 671B MoE model Cogito v2.1
Deep Cogito releases Cogito v2.1, a 671-billion-parameter mixture-of-experts open model aimed at complex reasoning and systems thinking.
Google DeepMind releases Gemini 3 Pro
The professional-tier base of Google's third-generation natively multimodal architecture, with more balanced performance across modalities.
Google DeepMind releases Gemini 3 Deep Think
A reasoning-focused derivative that incorporates search-tree mechanisms, trading longer compute time for accuracy on complex logic.
xAI releases Grok 4.1
A Grok 4 iteration that further strengthens high-throughput real-time data handling and extraction from complex social-network text.
Mario Zechner releases the Pi Coding Agent
Pi ships as a minimal, extensible terminal coding agent, with multi-model support, tool calling, session branching, and custom extensions.
OpenAI releases GPT-5.1
A maintenance update to GPT-5 that improves adherence to complex nested instructions and reduces hallucination on some tasks.