AIBlack Forest Labs' new image model does text-to-image, multi-reference editing and region placement; the distilled checkpoint makes a 1024 x 1024 image in about 2 seconds on an H100.
AIThe Swiss open model's July weights now come with a technical report: 4 trillion extra tokens for the 8B model, a thinking mode, intermediate checkpoints and evaluation cards in the EvalEval schema.
AIThe update to Ai2's climate emulator adds a downscaler that predicts several high-resolution fields at once and reproduces rare rain, wind and pressure extremes; models, notebooks and a demo are public.
AIThe reasoning-only model keeps its predecessor's architecture, extends context to 131K tokens and doubles the vocabulary; Liquid AI reports 184 tokens/s on an M5 Max and leads on tool-use benchmarks it selected.
AIAgentCore short-term memory is now billed for ingestion at $1.00 per GB, retrieval at $0.20 per GB and storage at $0.10 per GB-month, with a 12 KB minimum and 64 KB maximum billable size per event. Long-term memory pricing is unchanged.
AIAn arXiv preprint from East China Normal and Fudan universities reports that tokenizer-free Transformers beat subword ones as size grows, learn their own segmentation, and accept 3.4 times more speculative-decoding tokens.
AIMicrosoft released Microsoft-Decision-1, post-trained from Qwen3.5-9B, to score fixed options for routing, classification and agent control; all benchmark and speed figures are Microsoft's own.
AIAi2 replaced a priority scheduler on thousands of H100/B200/B300 GPUs with time budgets, hierarchical fair-share and a minimum-runtime contract; debug-job p90 wait fell from 2 hours to 30 seconds.
AIOpenAI's 8 October report: a Russia-origin network ran a Latin American research platform with unwitting staff, an Iran-origin one placed almost 100 articles under seven invented journalists.
AINVIDIA reports 535.4 of 600 on IOI 2026 in an unofficial live run and 30 of 42 on IMO 2026, graded by official IMO graders, from one model family and a four-step recipe.