Breakthrough 30B Model & Kimi-Linear-48B Released; Codex 5.3 Mimics Human Search
New AI Models & Releases
Experimental 30B Model with Subquadratic Attention Released: A new experimental AI model achieves 100 tok/s at 1M context and 76 tok/s at 10M context on a single GPU, offering significant memory and energy savings for long-context applications. The release includes the model, inference code, and a research paper.
Kimi-Linear-48B-A3B & Step3.5-Flash Models Now Available on llama.cpp: Two new models, Kimi-Linear-48B-A3B and Step3.5-Flash, have been released with GGUF quantizations, optimized for local inference via llama.cpp.
Model Updates & Benchmarks
Claude Opus 4.6 Outperforms Opus 4.5 in 3D VoxelBuild Benchmark: Benchmarks show Opus 4.6 delivers enhanced creativity and detail in 3D builds, rivaling OpenAI’s top model, though at a higher cost.
Opus 4.6 Now Available on Perplexity for Max Subscribers: Perplexity has integrated Opus 4.6 into its Model Council, allowing Max subscribers to compare it with other frontier models.
Codex 5.3 Exhibits Human-Like Search Behavior: Users report Codex 5.3 attempting multi-step problem-solving (e.g., searching for tools, installing libraries) instead of direct task execution, taking 35+ minutes for simple requests.
AI Industry & Partnerships
Goldman Sachs Adopts Anthropic’s Claude for Accounting & Compliance Automation: The financial giant is leveraging Claude to automate roles in accounting and compliance, aiming to improve efficiency and accuracy.
AI Policy & Advocacy
Call for European AI Investment to Boost Mistral’s Competitiveness: A proposal suggests Europe should establish a large government fund to support AI development, positioning Mistral AI as a key player despite limited resources compared to U.S. firms.