NVIDIA's new open 120B mixture-of-experts model fuses text, image, document, and video reasoning into one agent-ready backbone with 12B active parameters.
Sakana AI released The Agentic Scientist, an open-source scaffold that wraps coding agents like Claude Code and Codex into a rigorous end-to-end research loop.
A tiny Python harness lets Claude Code or Codex drive a real iPhone through macOS iPhone Mirroring, or any Android over adb, with no jailbreak and nothing installed on the phone.
A new reinforcement learning reward treats chain-of-thought as gradient descent, cutting tokens by about 50% while boosting accuracy on math, science, and code benchmarks.
A 315M-parameter Japanese embedding model built on ModernBERT-Ja tops JMTEB benchmarks while handling 8K context, outperforming OpenAI and billion-parameter rivals.
A 31B Gemma 4 fine-tune turns open language models into fast, calibrated decision engines for routing, extraction and classification in Polish and English.
IFM's new Arabic-first model family spans 0.9B to 375B parameters, with a flagship MoE that matches Claude Sonnet 5 on Arabic benchmarks and acts as an agent.
A training-free sparse attention method uses a handful of representative attention heads as proxies, delivering 10.3x attention speedup on long contexts.
Thinking Machines rolled out Dimuon, a LoRA-specific optimizer that beats AdamW on step count in supervised fine-tuning while leaving RL mostly unchanged.