Meituan released open weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and trained entirely on Chinese hardware without Nvidia GPUs (The Decoder). The model achieved 88.4 points on Terminal Bench, placing it just behind Claude Opus 5. Sample deployment runs on 194 Huawei Ascend chips, demonstrating an alternative inference path for large agent workloads independent of Nvidia infrastructure.
Google’s Gemini 4 Carbon took the top position on the MathArena benchmark for the first time in recent history, displacing OpenAI’s GPT-6.1 Sol (The Decoder). Anthropic’s Fable 5.2 also outperformed GPT-6.1 Sol, though both Google and Anthropic models carry significantly higher inference costs than the OpenAI baseline. Additionally, Cmpnd released dspy-wsp, a DSPy optimizer prototype for weight-space programming that enables frontier LLMs to edit smaller model weights in a sandbox without gradients at an estimated cost of $5 to $10 per run (MarkTechPost).