Cohere's Command A claims #3 open model spot (after DeepSeek and Gemma)

**Cohere's Command A** model has solidified its position on the LMArena leaderboard, featuring an open-weight **111B** parameter model with an unusually long **256K context window** and competitive pricing. **Mistral AI** released the lightweight, multilingual, and multimodal **Mistral AI Small 3.1** model, optimized for single RTX 4090 or Mac 32GB RAM setups, with strong performance on instruct and multimodal benchmarks. The new OCR model **SmolDocling** offers fast document reading with low VRAM usage, outperforming larger models like Qwen2.5VL. Discussions highlight the importance of system-level improvements over raw LLM advancements, and **MCBench** is recommended as a superior AI benchmark for evaluating model capabilities across code, aesthetics, and awareness.