Daily AI news, sourcedUpdates / Models
Updates Models · 505 items · newest issue first · page 11 of 26
AreaAllModelsAgents & codingApps & productsChips & infraResearchPricing & limitsBusiness & fundingPolicy & safetyOther
- LocalLLaMA user shares vllm-ascend progress running GLM-5.3-Flash on dual Ascend 310P cardsUnconfirmedr/LocalLLaMAOct 12Models
- Cactus Compute releases Whistle, a 16.9 MB on-device speech-to-text model for seven languages
- mlx-community posts a 6-bit Apple Silicon port of JetBrains Mellum2.1-12B-A2.5B-ThinkingUnconfirmedHugging Face @mlx-communityOct 11Models
- OpenAI will shut down the first GPT-4o API snapshot (gpt-4o-2024-05-13) on October 23, 2026OfficialOpenAI DeprecationsOct 11Models
- llama.cpp merges probabilistic MTP drafting with rejection-sampling verify for ~14% faster prose decode
- OrcaRouter releases OrcaCyber Zero 1.5, a gated 1M-context cybersecurity model, per MarkTechPost
- Qualcomm CEO Cristiano Amon says AI firms want 100B-parameter always-on phone models by 2028, per Sources
- Ollama v0.40.3-rc0 pauses background model compatibility migrations to cut high-volume overhead
- mlx-community publishes a Gemma 4 E2B OptiQ 4-bit decision model for Apple SiliconUnconfirmedHugging Face @mlx-communityOct 11Models
- llama.cpp PR #29416 adds MiniCPM-V 4.7 model supportUnconfirmedr/LocalLLaMAOct 11Models
- LocalLLaMA user says Qwen3.8-Flash-Next on Strata with pi beat Claude Code Opus 5.5 on an everyday-tasks bench on a 5090Unconfirmedr/LocalLLaMAOct 11Models
- LocalLLaMA user hits SWE-Bench LimitExceeded with Gemma 4’s agentic chat template vs Qwen3.8 27BUnconfirmedr/LocalLLaMAOct 11Models
- LocalLLaMA user runs Qwen3.8-Flash-Next-oQ4e-mtp on an M3 Max 64GB Mac with oMLXUnconfirmedr/LocalLLaMAOct 11Models
- MiniMax M3 runs on NVIDIA Vera Rubin NVL72 via vLLM; vLLM cites >7.8x GB200 throughput on AgentX
- Kling AI highlights creator tests of Kling 4.0 on surreal, true-to-life scenes
- User reports ~80 tok/s for Qwen3.6-35B-A3B on a 64GB Strix Halo mini PCUnconfirmedr/LocalLLaMAOct 11Models
- LocalLLaMA user benches three Qwen 3.8 quants on recent AtCoder problems across four thinking effortsUnconfirmedr/LocalLLaMAOct 11Models
- mlx-community publishes clef-flash OptiQ 4-bit for Apple SiliconUnconfirmedHugging Face @mlx-communityOct 11Models
- Custom CUDA megakernel claims 140 tok/s for Qwen3.8-27B on one RTX 3090Unconfirmedr/LocalLLaMAOct 11Models
- Community Qwen 3.8 27B Swift 1.5 W8A16 quant targets speed and quality on 48GB VRAMUnconfirmedr/LocalLLaMAOct 11Models