Daily AI news, sourcedUpdates / All
Updates Official · 943 items · newest issue first · page 3 of 48
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- MiniMax releases M3.1 Flash, says open weights will follow
- Inco AI releases a DFlash 2 speculative-decoding draft model for Qwen3.8-Flash-NextOfficialInco AI (Hugging Face)Oct 19Models
- Zhipu (Z.AI) paper reveals internal unified model GLM-OmniOfficialarXivOct 19Research
- Kapitsa, an open-source inference engine for Apple Silicon Macs from Nace.ai developers, supports only Qwen and Nemotron modelsOfficialKapitsa (GitHub)Oct 19Chips & infra
- Qwen Code 0.25.3 lets hosted agents launch read-only child agents and adds SSH, macOS and Kubernetes workspace runtimesOfficialQwenLMOct 19Agents & coding
- MiniMax-M3.1-Flash now listed in MiniMax API docsOfficialMiniMax API docsOct 19Models
- Fireworks AI adds NVIDIA Nemotron 3.5 Super VL as a training model with up to 262K context
- Baidu Qianfan adds DeepSeek-V4.1-flash-1017 to its model list with a 1M-token context
- Paper finds recursively self-distilling a safe model on its own safe outputs can make it progressively misalignedOfficialarXiv cs.LGOct 19Research
- Lean4AgenticRL trains 9B and 27B agent models with Lean4-verified process rewards, and the authors report 77.3% on TerminalBench-2.1
- Paper teaches a masked diffusion language model to distrust corrupted context, with a finetune that runs in under five hours on a laptop GPUOfficialarXiv cs.LGOct 19Research
- Paper finds physics-labeled prompts make an LLM a strong early guess for engineering simulator calibration, best used as a warm startOfficialarXiv cs.LGOct 19Research
- ImageEval 2026 entry detects multimodal hallucinations zero-shot with a 7B vision-language model and rotated answer choicesOfficialarXiv cs.CLOct 19Research
- AuditKIT, a source-available evaluation library from Lexsi Labs, puts benchmark, judge, runtime, tool-use and retrieval results in one format
- RealCLI benchmark tests CLI agents against a scripted user who changes instructions mid-task, and the best setup averages 49.2%
- Researchers train GPT-2-small models on 96 synthetic languages to test how language models learn action nominal constructions
- Paper replaces a small language model's learned embedding table with a frozen 16 MB vocabulary and reports matching or better benchmark scoresOfficialarXiv cs.CLOct 19Research
- SGLang 0.5.22 speeds up DeepSeek-V4.1 serving by up to 2.3x on H200, adds a 10x faster model loader and runs on Apple Silicon
- Epic Games' Unreal Engine skills plugin lets Claude Code control Unreal Editor over MCP
- OpenAI's desktop app now lets you edit previous messages within a thread, Tibo says