Meta: latest AI news
All Oossa news about Meta, newest first.
Known for: Meta AI, Llama, Muse
Follow via RSSToday
Models
llama.cpp release adds GQA support and new tokenizer features
The open‑source llama.cpp library now handles non‑standard GQA models and a BPE tokenizer, with expanded platform builds.
Models
llama.cpp b11818 adds typed input for reranking
The release lets the server accept typed content at its rerank endpoint and returns clearer errors for unsupported inputs.
Business
Google TPUs enter Meta’s production AI systems
Meta is running AI workloads on Google’s custom chips through PyTorch-native versions of two inference engines. The announcement gives few details on scale or availability.
Models
llama.cpp fixes SYCL Conv2D layout handling
Release b11815 fixes incorrect SYCL Conv2D results for certain tensor layouts and reports all 2,006 backend tests passing.
Models
llama.cpp release b11813 adds Tekken tokenizer and new binaries
The open‑source Llama.cpp library released version b11813 on Oct 22, 2026, swapping its default tokenizer for a Tekken‑based one and publishing fresh macOS, iOS, Linux and GPU binaries.
Models
llama.cpp adds new Q8_0 and i8mm kernels for Apple and Linux
The open‑source LLaMA inference library now includes faster quantized kernels for ARM SVE and i8mm, plus new binaries for macOS, iOS and Ubuntu.
Models
llama.cpp releases b11812 binaries for multiple platforms
The ggml‑org/llama.cpp project posted version b11812 on Oct 22 2026, adding pre‑built binaries for macOS, iOS and various Ubuntu builds.
Models
llama.cpp v b11810 adds new dequant switches and more binaries
The open‑source LLM runtime released binaries for macOS, iOS and several Ubuntu builds, and merges fp16/fp32 dequant switches.
Models
llama.cpp v b11806 adds LoRA weight support and new binaries
The open‑source LLM runtime released new LoRA prism.hadamard weight support and binary packages for macOS, iOS and Linux on Oct 22 2026.
Models
llama.cpp b11809 adds WebGPU support for Q2_0
The llama.cpp release adds Q2_0 row dequantization and matrix multiplication to its WebGPU backend. The release notes give no performance figures.
Models
llama.cpp release b11804 adds OpenCL argmax support
The open-source LLM runtime updates with F32‑to‑I32 argmax for OpenCL and new binaries for macOS, Linux and iOS.
Yesterday
Models
llama.cpp release adds Flint model and new binaries
The open‑source llama.cpp library released version b11803 on Oct 22 2026, introducing the FlintForCausalLM model and pre‑built binaries for macOS, iOS, and Linux.
Models
llama.cpp adds support for SenseNova U1 model
The open‑source llama.cpp library now runs SenseNova U1 dense understanding and vision models on CPU.
Models
llama.cpp adds fused top‑k mask operation in vb11800 release
The open‑source inference library now includes a new fused op for QSA masking, improving performance for certain models.
Tools & agents
Meta reports Muse AI app reaches 10 million users
Meta says 10M+ users use Muse, with a quarter shopping via the AI agent and average purchases over $200.
Business
Meta to issue non‑voting shares to fund AI spending
Meta plans new non‑voting stock, likely to sell more shares to pay for AI talent, chips and acquisitions.
Models
llama.cpp release b11799 adds SYCL q8 fast dequant support
The open‑source llama.cpp library version b11799, published on Oct 21 2026, adds a SYCL fast‑dequant routine and provides new binaries for macOS, iOS and Linux.
Policy & law
Meta leads big‑tech lobbying on AI antitrust rules
Meta, OpenAI and Anthropic filed lobbying reports showing they pushed Washington on AI information‑sharing rules in Q3 2026.
Models
llama.cpp b11797 release adds OpenCL IM2COL_3D support
The open‑source llama.cpp library released version b11797 on Oct 21, 2026, adding IM2COL_3D to its OpenCL backend and updating pre‑built binaries.
Models
llama.cpp release b11796 adds mtmd state object and new binaries
The open‑source LLM runtime llama.cpp posted version b11796 on Oct 21, 2026, adding a mtmd state object and offering fresh builds for macOS, iOS, Linux, Vulkan and CUDA.
Models
llama.cpp release b11795 adds server param forwarding
The open‑source llama.cpp library now forwards extra sampling parameters through its /v1/messages API and provides new binary builds for macOS, Linux and iOS.
Models
llama.cpp adds faster vocabulary loading in new pre‑release
The open‑source Llama.cpp library released a pre‑release on Oct 21 2026 that speeds up loading model vocabularies.
Models
llama.cpp version b11792 adds GPU stay‑on‑device support
The open‑source LLM runtime releases new binaries and a GPU optimisation that keeps data on the GPU longer.
Models
llama.cpp vb11790 adds neon‑vectorized repack for ARM devices
The open‑source Llama.cpp library released v b11790 on Oct 21 2026, adding neon‑vectorized and threaded repack support for ARM CPUs.
Models
llama.cpp 0.7.2 release fixes Metal GPU memory leaks
The open‑source Llama model library adds a Metal fix and ships new binaries for macOS, iOS and Linux.
Models
llama.cpp release b11788 adds SYCL fix and new binaries
The open‑source Llama.cpp library released version b11788 on Oct 21, 2026, fixing a SYCL crash and providing pre‑built binaries for macOS, iOS and Ubuntu.
Models
llama.cpp release adds OpenCL dp4a matrix multiply variants
The open-source llama.cpp library version b11789 adds new OpenCL kernels for faster matrix math and provides fresh binaries for macOS, iOS and Ubuntu.
Models
llama.cpp releases new binary builds for macOS, Linux and iOS
The open‑source LLM runtime adds updated binaries for Apple Silicon, Intel and several Linux GPU back‑ends on Oct 21 2026.
Models
llama.cpp adds faster CPU GELU kernels in v b11786
The open‑source LLM runtime updates its CPU math to compute GELU functions in full precision without a lookup table.
Models
llama.cpp adds per‑expert token scaling for MoE models
The open‑source llama.cpp library now scales MoE down‑projections per token, improving tensor‑parallel support.
