Daily AI news, sourcedUpdates / Models
Updates Models · 505 items · newest issue first · page 15 of 26
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- TII introduces Falcon-ASR, a 1.6B Arabic-focused speech model with 20.92% average Arabic WER
- jevman benchmark makes six decision models play Pac-Man, including Kev, Clef, GPT-6 Luna and LayaUnconfirmedr/LocalLLaMAOct 9Models
- UkisAI says its Swift Qwen models passed 2.2M downloads and offers early access plus free research computeUnconfirmedr/LocalLLaMAOct 9Models
- User shares an OpenAI-compatible OmniVoice TTS server with about 0.3s first audio on an RTX 3080Unconfirmedr/LocalLLaMAOct 9Models
- Local RTX 4090 benchmark compares Laya, Liquid D1, Clef Flash and Interfaze Lev decision models word-by-wordUnconfirmedr/LocalLLaMAOct 9Models
- Castmates developer shares on-device 3B roleplay finetune measurements from a native iPhone appUnconfirmedr/LocalLLaMAOct 9Models
- Local user attributes some Strata speed to an MTP head and ports one for llama.cpp GGUFsUnconfirmedr/LocalLLaMAOct 9Models
- Reddit user recommends niche Hugging Face finetune VeriLoop-E2 GGUF for coding workUnconfirmedr/LocalLLaMAOct 9Models
- OpenAI case study: Pollo AI uses GPT-5.6, GPT-6 Astra and GPT-Image-2.5 for campaign creativesOfficialOpenAI NewsOct 9Models
- JetBrains releases Mellum2.1, a 12B MoE open coding-agent model scoring 47 on SWE-bench Verified, per MarkTechPost
- Baidu Qianfan docs note DeepSeek thinking models support thinking_budget for output-length controlOfficialBaiduOct 9Models
- UniPat PaperBenchX finds GPT-6 Astra fully reproduces only 13.98% of 93 papers despite strong math claims, per QbitAIReportedQbitAIOct 9Models
- User says Halogen 0.17.2 holds about 45 tokens/s decode with Qwen 3.8 Flash Next on 128GB Strix HaloUnconfirmedr/LocalLLaMAOct 9Models
- Neuphonic open-sources NeuDecide, a 43 MB audio-to-tool model with a WASM browser demoUnconfirmedr/LocalLLaMAOct 9Models
- Developer trains Turkish TTS from scratch with the Drifting method on an RTX 5090Unconfirmedr/LocalLLaMAOct 9Models
- Local LLM user launches airbench.ai to share harness-and-model combination resultsUnconfirmedr/LocalLLaMAOct 9Models
- llama.cpp pull request speeds up Qwen 3.x prompt processing on CUDA by about 5%Unconfirmedr/LocalLLaMAOct 8Models
- Reddit user proposes a benchmark where models predict other models' outputs from their weights aloneUnconfirmedr/singularityOct 8Models
- Reddit user asks which local models are best for storytelling and world-building on an 8GB or 16GB GPUUnconfirmedr/LocalLLaMAOct 8Models
- Reddit user asks how to run Unsloth's Qwen3.8-Flash-Next MTP files with llama.cpp on two DGX SparksUnconfirmedr/LocalLLaMAOct 8Models