Headline News
AI & LLM · Featured
OpenAI ships GPT-6.2 Astra & Sol — frontier quality at a third of the cost
Astra sets new highs on Agents' Last Exam and Terminal-Bench-Science; Sol keeps the $2/$10 price for everyday agentic work. Third-party boards now rate Astra at 91.1, chasing Claude Fable 5.2's 93.2.
Hardware · Featured
AMD Instinct MI455X posts 21,333 tok/s on DeepSeek-R1, claiming 30% over Nvidia Vera Rubin
Fresh from SC26: AMD's Helios rack-scale system leads in inference throughput, and the new MI450X adds a 288GB HBM4 option to the MI400 family.
Browse the Universe
Latest Across All Decks
Inkling 2 Mini: open-weight laptop model
Thinking Machines ships a looped-MoE (26.4B total / 3.4B active) with Apache-2.0 weights on day one.
Nov 20–21, 2026llama.cpp hits 1.0 — stable library API
The backbone of local LLM inference finally guarantees C API/ABI stability with semantic versioning.
Nov 13, 2026Nvidia RTX Spark desktops out now
ASUS, Dell and MSI compact desktops with Grace + Blackwell and up to 128GB unified memory for 120B local models.
Nov 18, 2026Tencent Hy4 weights go open
770B MoE (49B active), 1M-token context, BF16/FP8 checkpoints on Hugging Face with a same-day AMA on r/LocalLLaMA.
Nov 20, 2026