Meta AI News

Meta delays its website-crawling business terms to 20 November as the UK CMA steps in

Meta pushed its new Commercial Terms, which let it crawl websites tied to business accounts, from 30 October to 20 November after publisher protests, a CMA letter and EU complaints.

24 Oct 2026News

Models8 min read

NVIDIA releases Nemotron 3.5 Bolt E4B, a 4GB model for local voice agents

NVIDIA's Nemotron 3.5 Bolt E4B 4GB is a Gemma 4-based small model trained with reinforcement learning only. It leads 4B peers on instruction following but loses some vision skill.

24 Oct 2026News

Industry7 min read

Meta buys AI analytics startup Sundial and brings its ex-Meta founders home

Meta has acquired Sundial, an AI data analytics startup founded by former Meta leaders Julie Zhuo and Chandra Narayanan, and is folding it into the data unit run by Alex Schultz.

24 Oct 2026News

Tools & Agents9 min read

Meta launches its Muse AI agent in Japan with Mercari, Money Forward and Ticket Pia

Japan is Muse's first market outside North America. The agent arrives on iOS, Android and muse.ai with local partners and a Klook booking flow, but without voice mode or Muse email yet.

23 Oct 2026News

Tools & Agents8 min read

1Password lets AI agents like Muse, Dots and Claude Duet sign in without seeing passwords

1Password for Agents fills logins for approved personal AI agents so the credential never reaches the model, starting with Meta, Anthropic, OpenAI, SpaceXSI and Town integrations.

22 Oct 2026News

Meta creates non-voting Class C shares to fund its AI push as Muse tops 10 million users

Meta will issue a non-voting Class C stock in January 2027 to fund AI spending, hiring and deals, while Zuckerberg said over 10 million people have set up Muse agents and a quarter use it daily.

21 Oct 2026News

Industry9 min read

Judge orders Meta to name everyone who approved downloading pirated books for Llama

A 19 October discovery order says Meta must identify every person behind its LibGen and Anna's Archive downloads, including who decided how to handle seeding over BitTorrent.

21 Oct 2026News

Models8 min read

Kakao releases Kanana-2 vision encoder weights in 426M and 98M sizes

Kakao's new open-weight Kanana-2 vision encoders lead Apple, Meta, Google and Qwen encoders on its own 42-benchmark average, with big gains on Korean OCR and documents, under a custom licence.

21 Oct 2026News

The TorchTPU banner: a golden torch with an orange flame on the left, the words Introducing TorchTPU, A Native PyTorch Backend for Google TPUs in the centre, and a chip labelled TPU with glowing circuit traces on the right, on a dark slate background.

Tools & Agents7 min read

Google open-sources TorchTPU so PyTorch runs natively on TPUs

Built with Meta, TorchTPU lets PyTorch models target Google's TPUs as a normal device, replaces the deprecated PyTorch/XLA bridge, and ships with TPU versions of vLLM and SGLang.

20 Oct 2026News

Tools & Agents9 min read

Meta Muse share links can be found through public DNS records, researchers say

Meta has paused Muse sharing after security firm Claroc found more than 6,800 share links in public DNS records, listed by conversation title, some exposing loan applications, revenue reviews and meeting notes.

20 Oct 2026News

Meta details Prometheus, the Ohio supercluster that trained Muse Wonder

Meta says the first phase of its gigawatt-scale Prometheus cluster in New Albany, Ohio, is complete, with hundreds of thousands of NVIDIA GPUs, over 85% pre-training goodput and a path to about a million GPUs.

20 Oct 2026News

Industry10 min read

WhatsApp help pages hid chatbot orders not to recommend privacy regulators

Redact found 102 hidden LLM instruction blocks on about 80 WhatsApp help pages, including orders not to recommend data protection authorities or cite GDPR. WhatsApp told WIRED a bug exposed them and removed the text.

19 Oct 2026News

AI-generated illustration of a wooden courtroom gavel resting on a tablet screen filled with abstract glowing network lines, with no readable text and no person.

OpenAI and Meta tell a federal evidence panel deepfakes will outrun reliable detection

At the Advisory Committee on Evidence Rules' 15 October 2026 Boston meeting, OpenAI and Meta panelists warned that synthetic audio and video are getting harder to authenticate, while a draft Rule 901(c) and an FJC survey still show few courtroom deepfake fights so far.

16 Oct 2026News

Industry12 min read

Meta delays its new business crawling terms to 20 November after publisher pushback

Meta has pushed its new Commercial Terms from 30 October to 20 November after publishers objected. The AI 'train' wording is already gone, but the crawling and test-credential rights remain, and robots.txt is still only in Meta's statements.

16 Oct 2026News

PrismML's off-white Bonsai mark, three stacked horizontal bars above an angled trunk shape, on a dark background on the left, beside a phone outline filled with rows of green, grey and dark dots on the right.

Models8 min read

PrismML releases Ternary Bonsai 2 in 2B, 4B and 9B sizes for phones and laptops

PrismML released three ternary vision-language models on 15 October 2026, Ternary Bonsai 2 2B, 4B and 9B, built from Qwen3.5 with every weight stored as -1, 0 or +1. The files are 0.43GB, 0.95GB and 2GB, the weights are Apache 2.0, and PrismML says they keep most of the full-precision models' benchmark scores.

16 Oct 2026News

Models9 min read

Meta releases Muse Spark Open, an 816B open-weight model that runs on two Mac Studios

Meta has published the weights of Muse Spark Open, an 816-billion-parameter multimodal model with 42 billion active parameters and a 1M-token context, built for coding and agents. Apple measured it at about 16 to 37 tokens per second on two to four M5 Ultra Mac Studios. The scores are Meta's own.

15 Oct 2026News

Research9 min read

Meta shares ten health studies doctors ran with its Muse AI models

Meta says a small group of physician-researchers used a Muse Spark 1.3 and Muse Wonder coding agent on about five million de-identified health records to produce ten observational studies in weeks. The team’s own report says the model was sometimes confidently wrong.

15 Oct 2026News

Tools & Agents8 min read

Hermes Agent v0.22.0 adds a free setup model, app connectors and one-click local models

Nous Research’s v0.22.0 ‘Threshold Release’ of its MIT-licensed Hermes Agent opens fresh desktop installs on a guided setup chat that runs on a free trial model, adds hosted app connectors without an account, installs llama.cpp for local models in one click, and moves installs to tested stable releases.

Facebook’s white lowercase f inside a solid blue circle, centred on a plain dark navy background.

Facebook adds Your algorithm so people can steer Feed and Reels in their own words

Meta brought Instagram’s Your algorithm tool to Facebook on 13 October, letting people add topics and explain in plain language what they want more or less of in Feed and Reels, with large language models now feeding the ranking systems.

Tools & Agents7 min read

Hugging Face puts llama.cpp in a Windows 11 tray app called Llama Windows

Llama Windows is a free, open-source WinUI 3 app that installs llama.cpp without admin rights, downloads GGUF models from the Hugging Face Hub in one click, and exposes an OpenAI-compatible local server at localhost:9931, with day-one support for Windows on Arm and NVIDIA RTX Spark.

The Palace of Westminster lit up at dusk across the River Thames, with Victoria Tower on the left, Elizabeth Tower and its clock face on the right, and the floodlit facade reflected in the water.

Industry9 min read

OpenAI and Anthropic back binding AI safety rules at a UK Parliament hearing

At a Commons committee hearing on 13 October, Anthropic said AI safety rules should be binding and OpenAI backed mandatory incident reporting, while Google DeepMind and Meta would not support new legislation. UK officials admitted that model testing by the AI Security Institute is voluntary.

Dozens of coils of yellow fibre-optic cable hang in loops from racks along a server-room wall, with grey and blue cables running beneath them.

Meta unveils liquid-cooled 102.4T switches and a new GPU fabric at OCP 2026

At the Open Compute Project summit, Meta showed IcePack, its first liquid-cooled network rack, a fabric that gives each GPU eight separate network paths, an Ethernet switch built for its MTIA chips, and open-source MetaRoCE transport code.

13 Oct 2026News

A long, windowless grey and pale-green data center hall beside a gravel path and grass field in Luleå, Sweden, with a smaller brown service building and fenced electrical equipment on the right under a cloudy sky.

Models10 min read

Meta launches Muse Wonder and Muse Wonder Full frontier models for agents and coding

Meta’s Muse Wonder is live on the Meta Model API at $4 and $20 per million tokens, with a 1M-token context; the stronger Muse Wonder Full is limited to Muse Code subscribers. Meta’s own scorecard puts Full ahead of GPT-6 Astra and Claude on agent and long-horizon coding tests, but behind Claude Opus 5.5 on several others.

12 Oct 2026News

The concrete entrance of the San Jose McEnery Convention Center, decorated with a large red and black geometric mural, with fountains and stone spheres on the plaza in front under a blue sky.

Industry10 min read

NVIDIA opens Vera Rubin rack designs at OCP and plans USFP optics with Cisco

At the OCP Global Summit NVIDIA contributed Vera Rubin rack and tray designs through MGX 1.2, published a Vera CPU rack spec for agents, co-authored an open data center spec with AMD, Google and Meta, and unveiled USFP, a planned 400 Gb/s-per-lane pluggable optic built with Cisco.

12 Oct 2026News

Backlit chiclet keyboard of a mid-2012 MacBook Pro filling the frame. No people appear.

Tools & Agents6 min read

Humanizer ships app 0.3.3 for its local 12B AI-draft rewriter

On 10 October 2026 the humanizer project released macOS and Windows app 0.3.3 for its Apache-2.0 12B rewriter that turns AI drafts into more human-sounding prose while aiming to keep numbers, names, dates, and quotes. The model fine-tunes google/gemma-4-12B and runs offline after download; detector and fact-fidelity scores are the authors’ own.

11 Oct 2026News

A Google Tensor Processing Unit 3.0 accelerator board on a dark surface. No people appear.

Models6 min read

Kisoku 1.6B open-weights a solo from-scratch TPU Research Cloud pretrain

Joseph Rodriguez’s Kisoku 1.6B (Hugging Face created 7 October) is a from-scratch 1.6B base model trained solo on a TRC TPU v4-32. Report draft 4 claims a 5–5 split with Llama 3.2 1B on a shared harness with ~0.5T tokens. Scores are the author’s.

10 Oct 2026News

An open laptop on a wooden desk beside a keyboard, mouse, and notebook. No people appear.

Tools & Agents6 min read

Base Compute opens Superfluid for multi-agent local LLM serving

Base Compute’s 8 October Hugging Face post introduces Superfluid, an Apache-2.0 local LLM server aimed at several agents sharing one machine. It speaks OpenAI, Anthropic, and Ollama APIs over llama.cpp, MLX, and BaseRT. We did not install it.

10 Oct 2026News

A black PUMA robotic arm with NASA lettering on a black background, hovering over a white dish of block shapes. No people appear.

Research8 min read

Meta FAIR’s RoboJEPA repo is live with CC BY-NC-SA weights

A 7 October arXiv preprint from Meta FAIR and Mila fits a second-order compute law for an 8B latent JEPA predictor on 12 robot embodiments. Table 3’s Franka scores remain the authors’. On 10 October the named github.com/facebookresearch/robo_jepa repository resolves with README, src/, configs/, and CC BY-NC-SA 4.0 license text, plus dl.fbaipublicfiles.com weight URLs. We did not download multi-gigabyte checkpoints or run a robot.

9 Oct 2026Analysis

Open Framework Laptop 13 chassis showing the motherboard, fan heat pipe, Lexar 7900 NVMe drive, wireless module and Framework battery. No people appear.

Models7 min read

Underdog releases Saluki 27B, a 7.89 GB 2-bit Qwen3.8 GGUF

The 7 October Underdog page and Hugging Face card ship an Apache-2.0 IQ2-mix build of Qwen3.8-27B for stock llama.cpp. Underdog Bench scores 88 of 120 against 84 for the full model. Math retention is about 82–85%. Those scores are Underdog’s.

9 Oct 2026News

A seven-storey brick university building with regular window bays, a roof terrace and three parked cars on a lawn in front.

Research7 min read

Nature publishes Ai2’s Bolmo: retrofit Olmo, Qwen and Llama to bytes

A 7 October Nature paper describes byteification: turning an existing subword LLM into a byte-level model with 49.1 billion extra tokens. Ai2 points to Bolmo, Bwen 8B and Blama 8B. Hugging Face created the Bwen and Blama cards in August. The scores are the authors’.

The entrance road to Facebook’s Menlo Park campus, with a low office building, blossoming trees and a white sign showing a thumbs-up icon and 1 Hacker Way.

Industry7 min read

Meta adds an LLM to flag ads that signpost off-platform child-exploitation material

Meta’s 7 October post describes a new language model for “signposting” ads, plus destination checks, AI ad sweeps and a red-teaming agent. The January–June 2026 action counts are Meta’s. We did not test the systems or review ads.

The National Institutes of Health Building 1, a brick building with a white-columned portico and snow-covered trees, photographed in winter.

Industry7 min read

Biohub, DOE and NIH expand the Virtual Biology Initiative to a $1.8 billion data stack

Biohub’s 7 October post adds multi-year DOE money, NIH dataset coordination, and $300 million from DeepMind, Isomorphic Labs and Meta. The $1.8 billion total mixes new pledges with Biohub’s April commitment and prior NIH investment.

Sierra co-founder Bret Taylor holding a microphone and gesturing while speaking on a TechCrunch Disrupt stage.

Tools & Agents6 min read

Sierra publishes Personal Agent Protocol draft 0.1, Poppy. OpenAI is among 35 new design partners

The 9 October Sierra post publishes a draft of Personal Agent Protocol, called Poppy, and lists 35 more design partners including OpenAI. The v0.1 text is at personalagentprotocol.org under Apache 2.0. A reference implementation is still promised.