Daily AI news, sourcedUpdates / Models
Updates Models · Reported · 38 items · newest issue first · page 1 of 2
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- Anthropic is preparing a model that better infers what users intend from limited information, per The InformationReportedThe InformationOct 20Models
- NVIDIA's Structured Data Models library makes zero-training predictions on tables and relational databases, per MarkTechPostReportedMarkTechPostOct 18Models
- UBTech says its Thinker-WM embodied world model tops WorldOlympiad, EWMBench and WorldArena, per QbitAIReportedQbitAIOct 17Models
- Intel CEO Lip-Bu Tan says Chinese frontier AI models are cheaper per token and being adopted 'very fast', per Reuters
- Vercel data shows TypeSafe AI's Jev decision model reached 35% of paid teams in two weeks, per TechCrunch
- Elorian launches Glimpse, a visual-reasoning model with a mid-reasoning crop-and-measure scratch space, per TechCrunchReportedTechCrunchOct 17Models
- ARC Prize verifies Gemini 4 Argon at 97.5% on ARC-AGI-3 with its provider adapter harness but 49.8% on the standard harness, per OfficeChaiReportedOfficeChaiOct 16Models
- Artemis Security launches Orion-1, a cyber-defense model that it says beats Claude Opus 5.5 and GPT-6 Sol on its own benchmark, per The Decoder
- Microsoft researchers publish Probabilistic Bias Correction in Nature, an open-source CPU-only tool that improves two-to-six-week weather forecasts, per The VergeReportedThe VergeOct 15Models
- Zhipu quietly tests GLM-5.3-Highspeed in Zcode, with a user clocking 490 tokens/s, per QbitAI
- National Hurricane Center leaned on Google DeepMind's WeatherNext 3 for Hurricane Isaias, per CNN
- OrcaRouter releases OrcaCyber Zero 1.5, a gated 1M-context cybersecurity model, per MarkTechPost
- Nace.AI open-sources Drex 1.5, a 9B decision model that scores options in one pass, per MarkTechPostReportedMarkTechPostOct 10Models
- Google employees are testing an internal Gemini 4 version called Carbon that feels like Opus 5.5 for coding, per Business InsiderReportedBusiness InsiderOct 10Models
- Robot Era's VPP2 world action model tops the RoboDojo simulation leaderboard ahead of GPT-6-Astra and pi0.5, per QbitAI
- Underdog launches Saluki 27B, a ~7.89GB Qwen3.8-27B quant retaining ~96% average score, per KuCoinReportedKuCoinOct 9Models
- JetBrains releases Mellum2.1, a 12B MoE open coding-agent model scoring 47 on SWE-bench Verified, per MarkTechPost
- UniPat PaperBenchX finds GPT-6 Astra fully reproduces only 13.98% of 93 papers despite strong math claims, per QbitAIReportedQbitAIOct 9Models
- Kaiming He's team solves ARC puzzles with a vision model pretrained on everyday ImageNet photos, per QbitAIReportedQbitAIOct 8Models
- Coding-tuned AI models have stopped talking like people, per QbitAIReportedQbitAIOct 7Models