Daily AI news, sourcedUpdates / Models
Updates Models · 505 items · newest issue first · page 6 of 26
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- Stanford's Hazy Research releases NCM, a non-autoregressive language model it says decodes up to 31x faster at matching quality
- Paper Instruments releases Ultramarine 1 Flash, an RL post-train of GLM-5.3 Flash for knowledge work, via API with weights promisedOfficialPaper InstrumentsOct 15Models
- Fish Audio releases Drama 3, a 35B open-weight mixture-of-experts text-to-speech model directed in plain language across 83 languagesOfficialFish AudioOct 15Models
- Datalab releases OmniParseBench, an open document-parsing benchmark of 16,288 pass/fail tests across 92 languagesStatementDatalabOct 15Models
- Anthropic releases Salesforce in Claude in beta, a plugin with 37 skills for account research, call prep, pipeline review and CRM updatesOfficialAnthropicOct 15Models
- Vercel's AI Gateway adds Microsoft's MAI-Image-2.6 and MAI-Image-2.6-Flash for image generation and editing
- Google DeepMind and GovTech Singapore fine-tune Gemma into an on-device scam detection model that keeps conversations private
- LightSeek's TokenSpeed serves Qwen3.8-Flash-Next at 685 output tokens per second per user on four NVIDIA GB300 GPUsOfficialLightSeek FoundationOct 15Models
- Perplexity adds Ultrafast versions of GPT-6 Astra and GPT-6.1 Sol to Perplexity Computer for Max subscribersOfficialX @perplexity_aiOct 15Models
- Cohere shows its Rerank models playing Mario, with Rerank 5 trying to beat Rerank 4's score
- LocalLLaMA user compares Qwen3.8-27B and Qwen3.8-Flash-Next on one threat-intelligence app build and would build on the 27B's codeUnconfirmedr/LocalLLaMAOct 15Models
- Sabi-Noema, an open-source RL fine-tune of Qwen3.8-27B, scores 76% on Terminal-Bench 2.1 while using about 2.3x fewer output tokens than the base modelOfficialHugging Face @sabicapOct 15Models
- Microsoft Foundry updates GPT-chat-latest with GPT-6-level intelligence, GenUI interactive answers and interleaved-thinking outputOfficialMicrosoftOct 15Models
- Microsoft Research releases UserRepGen-4B, an MIT-licensed model that turns conversations into task and behavior descriptions for user simulatorsOfficialHugging Face @microsoftOct 15Models
- xAI releases Grok Voice 3.0, which tops Artificial Analysis's speech-to-speech index, in its Voice Agent API at Grok Voice 2.0's $0.08-a-minute priceOfficialxAIOct 15Models
- Budget-matched 15M decision models show a linear-recurrent backbone running 2.37x faster than attention at 8K context, with accuracy near chanceOfficialGitHub @stube123890-hueOct 15Models
- Lech Mazur's creative story-writing benchmark ranks Muse Wonder (high) 14th of 59 models, ahead of Gemini 4 Argon but well behind Claude Fable 5.1Unconfirmedr/singularityOct 15Models
- TheoPhysEval: GPT-6 Astra in Codex solves 49 of 50 unpublished research-level theoretical physics problems at under $3 eachOfficialarXivOct 15Models
- Telperion releases Qwen3.8-27B-MLA and Qwen3.8-27B-GLA-g2, training-free attention retrofits that halve the model's KV cacheOfficialHugging Face @TelperionAIOct 15Models
- Zellige AI releases Tessera-27B-B-CJK, a Qwen3.8-27B coding-agent fine-tune that thinks in compressed CJK shorthand to use about half the thinking tokensOfficialHugging Face @ZelligeAIOct 15Models