A roughly $6B deal to acquire or heavily invest in a model factory whose models outperform reference benchmarks, flagged by Latent Space as the emerging AI-native company landscape takes shape.
At the World Robot Conference, Unitree's founder said robots still struggle to generalize while its shares fell 18.7% after an IPO-day surge; Keenon said it has shipped over 100,000 robots and expects 150,000 by end of next year.
A 280B-total/16B-active MoE with 512K context and recursive self-critique, featuring a test-time agentic value model (TEMPO) and a perfect 42/42 in a math-proof harness at IMO 2026. Free preview on OpenRouter.
A project-grounded, verification-gated agent harness that only declares tasks complete when external checks pass. Record 72.6% mean strict verified pass rate across seven models and a dynamic-behavior score of 52.2 vs 22.4-31.4 for baselines on live PLC runtime traces.
One LLM writes executable Gym-style environments while another solves them. Scaling to 30B parameters improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out benchmarks, and lifts tool-use by +5.7 on BFCL-v4 and +13.9 on ACEBench-Agent.
LLMs run a football club for 20 in-game years with ~340-400 decision stops and 26 tools; all 15 frontier models finish where scripted baselines die, and claude-fable-5 tops both solo and Arena tracks with the title rotating among ten models.
Giving skill-naming tokens an action-local advantage in a 16-candidate slate lifts a 9B policy from 40.8% to 53.2% trial success across five agentic benchmarks, while cutting misleading skill exposure by two thirds.
Examining a large set of public post-training trajectories, researchers find agents fix their initial training strategy early and then only adjust locally. Experience-driven scaffolds gave +12.6 on GSM8K and +40.8 on HumanEval, while extra inference compute did little to enable mid-execution shifts.
Together AI showcases Navigator, which reportedly achieves 2-5x better cost and speed than closed models on web tasks via ~80% prefix cache hit rates, speculative decoding, and heavy SFT+RL post-training of Qwen models.
Philipp Schmid's GitHub hub covers 16 Gemma variants with setup guides for Ollama/vLLM/LiteRT, fine-tuning recipes via Unsloth, Tunix and MLX, plus community projects.
Very small, scalar-valued Python-with-loops model specs could anchor the extreme efficiency end of open-weights development, Karpathy argues — anything larger being a compilation problem. A design note, not a release.
Released via an anonymous third-party provider, it offers a 1M-token context window, text/image/video input and tool calling, positioned for coding and long-horizon agentic work.
Agent applications can trade directly via MCP and other developer tooling, though Binance warns that agent 'keep-safes' remain largely the trader's responsibility. Rival exchanges are moving similarly.
The third open-sourced entry in Codacy's Skills series lets agents discover code paths and auto-configure test coverage, keeping instrumentation and assertions in sync with agent-generated code.
An open-source, AI-first composable package ecosystem launched on Show HN, aimed at letting agents compose and manage packages as modular building blocks.
An API gateway compatible with both the OpenAI and Anthropic Claude Messages formats, letting one endpoint serve tooling built for either protocol.
Source:
