AI news
New open models, runtime releases, and announcements from AI labs, a few lines each. Toskar Pro and Core show the same news, with Install for models in Toskar's catalog.
Updated every morning. The summaries are written by AI, a local model running in Toskar, and checked automatically; titles and links come from each source. Follow along in a feed reader with the RSS feed.
Friday, October 23, 2026
AnnouncementHugging Face
Does Your Model Run on AWS Trainium? Look It UpHugging Face has a new post on whether your AI models run on AWS Trainium chips. It's worth checking to see compatibility and potential performance gains.
Thursday, October 22, 2026
AnnouncementHugging Face
Ensemble forecasting with IBM Granite time-series modelsIBM Granite time-series models combine multiple forecasts to improve accuracy. This ensemble approach helps AI model runners get more reliable predictions.
Runtimellama.cpp on GitHub
llama.cpp b11826: add MiniCPM-OCR 1.0 supportllama.cpp, used in many local AI apps, has been updated to support MiniCPM-OCR 1.0. This allows for easier deployment of OCR models on various devices.
Wednesday, October 21, 2026
AnnouncementHugging Face
Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and FasterGranite Speech 5.0 improves accuracy and speed in transcribing, translating, and attributing audio. This update is useful for anyone processing large volumes of audio data on their AI models.
AnnouncementMistral AI
Mistral Advances AI for Engineering with NVIDIAMistral AI is working with NVIDIA to create better AI tools for engineering tasks.
ModelHugging Face
flint (39.8B) from RecursiveSI, in GGUFFlint, a 39.8B parameter model, excels in coding tasks like fixing bugs and writing tests. It scores 89.7% on SWE-bench Pro v2, making it a strong choice for developers.
Runtimellama.cpp on GitHub
llama.cpp b11773: add H2O-Lightning-4B decision modelThe new H2O-Lightning-4B decision model in llama.cpp b11773 can answer typed decision questions and uses up to 255 labels. This update allows local AI apps to run more sophisticated decision-making models.
Runtimellama.cpp on GitHub
llama.cpp b11801: Add support for SenseNova U1 modelllama.cpp version b11801 now supports the SenseNova U1 model, adding dense understanding support and vision encoder conversion. This update is important for users running local AI models as it expands model compatibility and functionality.
Runtimellama.cpp on GitHub
llama.cpp b11803: add Flint model (FlintForCausalLM)The Flint model (FlintForCausalLM) has been added to llama.cpp, allowing local AI apps to use this new model. This update is important for anyone running AI models locally as it expands their model options.
Tuesday, October 20, 2026
AnnouncementAi2
FairStress: A deeper evaluation of fairness in LLM decision-makingFairStress highlights how LLMs may treat different demographics differently, even when their decisions seem fair. This matters because it shows we need to carefully examine how these models explain their choices.
AnnouncementMistral AI
Voxtral 3Mistral AI has released Voxtral 3, a voice agent SDK supporting 50+ languages with real-time speech transcription and emotional text-to-speech. This update is useful for developers integrating voice capabilities into their applications.