Meta Releases Its First Open-Weight AI Model in More Than a Year

Meta Superintelligence Labs released Muse Glimmer on Monday, a 30-billion-parameter model that is Meta’s first open-weight large language model in more than a year. Distributed under an Apache 2.0 license, Glimmer is distilled from Meta’s larger, still-proprietary Muse Spark model and is optimized for always-on local agent workflows — tool calling, local coding, and LLM-as-a-judge evaluation — with multimodal input across more than 100 languages and a size Meta says fits on a Mac or PC with a single consumer GPU after quantization to roughly 4-bit precision.

The release accompanied a letter from CEO Mark Zuckerberg arguing that superintelligence should be distributed rather than concentrated, and Meta said it will open the weights of its flagship Muse Spark 1.2 within weeks. The Register reports that Glimmer beats Google’s Gemma 4 31B in most benchmarks and trades blows with Alibaba’s Qwen 3.6-27B, figures that remain self-reported; the weights are on Hugging Face, with integrations for Ollama, llama.cpp, MLX, and ExecuTorch expected in the coming days.

Featured source: The Register , Meta AI Research .

Other Stories

OpenAI’s AI-Generated Math Results Draw a Credit Dispute

The Verge reported Tuesday that OpenAI’s announcement of ten advances in mathematics and theoretical computer science — generated by its unreleased Astra model, prepared into papers by humans, and verified with Lean certificates — has produced a dispute over credit. Cambridge mathematician Francesco Fournier-Facio and Alfréd Rényi Institute researcher Gábor Kun said OpenAI’s original post understated the prior work of Kun and Andreas Thom that its construction establishing the existence of non-sofic groups builds on; after contact, OpenAI revised the language “to better reflect the prior research these results build upon.” The feature also quotes Fields Medalist James Maynard, who said he is “soul searching” about his field, as researchers weigh the funding, credit, and training consequences of AI-accelerated discovery. OpenAI says generating the ten solutions would have cost roughly $2,000 in tokens at current API prices, and the results are not yet peer-reviewed.

Filed from: The Verge , OpenAI .

OpenAI Adds a Dedicated Cyber Model to Its Daybreak Service

OpenAI expanded Daybreak, its cybersecurity service, into two tiers and introduced GPT-5.6-Cyber, a model trained for authorized security work. The Blue tier offers frontier models including GPT-5.6 Sol with safeguards calibrated for defensive tasks such as vulnerability discovery, malware analysis, and incident response; the Red tier adds purpose-trained models including GPT-5.6-Cyber for authorized vulnerability research and exploit validation, restricted to approved defenders with additional monitoring. OpenAI said it used the model in research that found previously unknown vulnerabilities in popular open-source software, including Chrome’s V8 engine, and access is reportedly limited to trusted partners such as Accenture, IBM, CrowdStrike, and Cloudflare.

Filed from: OpenAI , TechCrunch .

Sila Lands a $1.4 Billion Pentagon Loan to Expand US Silicon-Anode Output

The U.S. Department of Defense on Friday announced a $1.4 billion loan to battery-materials startup Sila to expand production of its silicon-carbon anode material at its Moses Lake, Washington factory, which began operating in September and currently makes about 2 gigawatt-hours of anode material a year. Silicon anodes can store roughly 20 to 40 percent more electricity than the graphite used in most lithium-ion batteries, and the graphite supply chain is dominated by Chinese companies. Sila, which has deals with Mercedes and Panasonic and raised $300 million in July, plans a fivefold expansion that it says would produce material for more than 100,000 electric vehicles. The Pentagon simultaneously announced loans to Sunrise Energy Metals ($400 million for scandium), Niron Magnetics ($150 million for rare-earth-free magnets), and an $85 million equity investment in Strategic Bauxite; the loan details come from TechCrunch’s reporting.

Filed from: TechCrunch .

North Korean Spy Group Kimsuky Is Running Local LLMs, Researchers Find

South Korean security firm Genians said Monday that a months-long investigation found Kimsuky — a cyber-espionage group tied to North Korea’s Reconnaissance General Bureau — operating local LLM environments built with Ollama, GPT4All, and Msty, using retrieval-augmented generation for local document search, and collecting libraries to integrate commercial AI services such as OpenAI and Azure AI packages into custom tools. Genians says the local setup keeps conversation data out of cloud services, and that AI-polished phishing lures targeting virtual assets and finance use natural language and business-like formatting to induce victims to open malicious LNK files. The researchers found no evidence Kimsuky is training its own models; the collected components point instead to building AI-augmented tooling for malware development and attack operations.

Filed from: The Register , Genians .

Amazon Backs a Texas Gas Plant That May Become the Largest US Source of Climate Pollution

Amazon confirmed it is investing in a natural-gas power plant in Pecos County, Texas, to supply electricity for a planned data center that will operate off the grid — the company’s first major investment in that model, according to Ars Technica. The New York Times reported the plant “could become the largest single source of climate pollution in the United States.” Amazon said the arrangement keeps the facility from straining the grid or raising electricity costs for neighbors, but did not answer Ars’s questions about the plant or the backlash. The project follows public fights over gas turbines powering xAI’s Colossus data center, and Ars reports that Anthropic, Google, Meta, Microsoft, OpenAI, and Oracle have also pursued self-supplied power projects.

Filed from: Ars Technica .

OVH Cloud Warns of Price Hikes Up to 87% as Memory Costs Soar

OVH founder and CEO Octave Klaba said the French cloud provider will raise server rental prices by up to 87 percent, attributing the increases to component costs that he says rose sharply as hyperscalers bought up supply during the AI boom. Klaba said the price OVH pays for memory rose sixfold in the year to June 2026 and is projected to reach 12 times current levels next year, with NVMe drives up 7 times and CPUs, motherboards, and network cards up 15 to 20 percent. The steepest rise applies to new 2026-edition gaming servers; other recent servers will cost 40 to 59 percent more from September, while renewals of 2024-era equipment rise less. OVH will also decouple storage and IP addresses into separate billing lines from October 1 and is dropping its 1-, 6-, and 24-month price-lock plans.

Filed from: The Register .

Virginia Orders Data Centers to Pay for Dedicated Grid Infrastructure

Virginia’s State Corporation Commission has ordered data centers to pay for transmission infrastructure used exclusively by their facilities, so utilities can no longer spread those dedicated costs across their wider customer base. The state operates at least 570 data centers, the most of any US state, and electricity prices in the region have risen sharply alongside the expansion — PJM Interconnection raised capacity prices by 76 percent in recent changes that the market watchdog Monitoring Analytics largely attributes to data center demand. State officials said the requirement should prevent ordinary electricity customers from carrying costs tied to large data center developments.

Filed from: TechRadar .

Anthropic Pledges EU Watermarking of Claude Output

Anthropic said it will embed watermarks in text and files generated by future Claude models it launches in the EU, citing the bloc’s AI Act transparency rules. “Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported,” the company said in a help document published Monday, adding that marking will apply wherever Claude is offered worldwide and that it is retrofitting already-released models during the EU’s transition period. The marks would cover output from Claude on the API platform, Claude Code, Claude Cowork, and Claude Tag, including third-party providers such as AWS, Google Cloud, and Microsoft Foundry.

Filed from: The Register .

OpenAI Completed a $7 Billion Employee Tender Offer

Bloomberg reported that OpenAI bought back $7 billion worth of shares from employees, a liquidity move that valued the privately held lab at $852 billion — the same as its most recent fundraising round in March, which added $122 billion to its war chest. OpenAI filed confidentially with the SEC in June to prepare for a possible IPO later this year and did not respond to a request for comment. TechCrunch notes that a tender at a flat valuation suggests an IPO may not be imminent, as the Wall Street Journal reported in April that OpenAI missed internal financial goals; the buyback is reported rather than confirmed by OpenAI.

Filed from: TechCrunch .

YouTube Doubles Watch-Time Thresholds for New Creators

YouTube said Monday that creators joining its monetization programs will need 8,000 qualified watch hours over the past year or 20 million qualified Shorts views in the last 90 days — double the previous 4,000 hours and 10 million Shorts views, with the subscriber requirement unchanged at 1,000. The change takes effect February 1 and will not affect creators already in the YouTube Partner Program. YouTube also raised the Shorts Creators Pool bar to a maintained 10 million Shorts views per 90 days for short-form revenue and framed the changes as keeping pace with growth, citing over 200 billion daily Shorts views and over a billion hours of daily TV watch time.

Filed from: TechCrunch .

Appeals Court Lets Thousands of Social Media Addiction Lawsuits Proceed

The 9th US Circuit Court of Appeals denied Meta, TikTok, Snapchat, and Google’s attempt to dismiss thousands of consolidated lawsuits claiming they designed products to be addictive to minors, rejecting for now their argument that Section 230 shields them from failure-to-warn claims. The court said the appeal came too soon, as that type of challenge typically arrives after trial, according to TechCrunch’s report on Reuters coverage. The cases, filed by private individuals, state and local governments, and school districts, will proceed as a consolidated federal suit; Meta has already lost two jury trials over similar child-safety claims.

Filed from: TechCrunch .

Amazon Data Center in Gilroy Advances Under Decades-Old Zoning Rules

Amazon Web Services is building a data center on a 56-acre parcel of farmland in Gilroy, California, approved under a 45-year-old zoning rule intended for industrial development along a freeway. Residents quoted by the Wall Street Journal say the project bypassed public planning scrutiny and that many first learned of it after construction had begun. AWS’s vice president of economic development said the project “followed standard approval processes” and that the company has since met with residents and hosted an open house; Amazon has donated $1 million toward a new city fire truck, and the mayor framed the project as part of adapting to changing times.

Filed from: TechRadar .

From the Community

Exploiting System Management Mode With a Very Long Interrupt

Security researcher Christopher Domas published a proof-of-concept showing how to break System Management Mode’s all-cores rendezvous using a single machine instruction that stalls for over a second, holding one core outside SMM while others enter. This desynchronization could enable software-only exploitation of dormant SMM TOCTOU vulnerabilities, which previously required physical access or a malicious DMA device. The PoC is tuned for a Zen 3 Ryzen 7 5800H, and the researcher notes there is no clear fix: keeping the timeout allows the attack, removing it risks hangs, and increasing it hurts performance on many-core platforms.

Filed from: GitHub repository , Hacker News discussion .

Needle 2: A 14MB Agentic LLM for Tiny Devices

Cactus Compute released Needle 2, an open 45M-parameter model for tool calling and structured extraction that fits in a 14MB binary and runs a full session in 28MB of RAM, targeting sub-$200 phones, wearables, microcontrollers, and robots. It claims 500 tokens/sec on a Raspberry Pi 5 and 300–700 tokens/sec on budget phones, using 2-bit quantization trained from pretraining and a co-designed C++ engine. On tool-calling benchmarks it trades wins with larger models like FunctionGemma 270M and Apple FM while being 5 to 70 times smaller, though the company notes benchmarks compare against f16 baselines and the model is specialized for agentic tool calling.

Filed from: Cactus Compute , Hacker News discussion .

H3-metal: Native MiniMax-H3 Inference for Apple Silicon

Antirez has released h3.c, a native Metal inference engine for MiniMax-H3 that runs entirely on Apple Silicon, supporting prompt-to-video/audio generation, first/last-frame conditioning, and ordered references. It includes SSD streaming to reduce memory usage from about 36.5 GiB to 2.0 GiB, and various speed/quality presets, with the author reporting a four-pass denoise taking about 3.5 seconds on M5 Max versus 26.4 seconds for a 29-pass reference. The project is under active development and performance figures come from the author’s own benchmarks.

Filed from: GitHub repository , Hacker News discussion .

Rust Portable SIMD Now Runs on the GPU

VectorWare announced that Rust’s portable SIMD (core::simd) can now target GPUs, mapping a Simd<T, N> vector directly onto a warp’s lanes, so the same source code runs on CPU and GPU with operations like elementwise arithmetic, reductions, and shuffles lowering to warp instructions. The company built a type-level IR to handle lane-width mismatches and a reference interpreter for differential testing. The implementation works on NVIDIA today, with AMD and Vulkan subgroups as future targets, though portable SIMD remains unstable in Rust and zero-cost mapping only holds when the vector width matches the warp width.

Filed from: VectorWare , Hacker News discussion .

As AI Eats the Web, the Internet’s Collective Memory Is Disappearing

An essay in The Walrus by Vass Bednar argues that AI-powered search is accelerating the decay of the web’s archival infrastructure, citing Google’s AI summaries inventing facts, the deletion of FiveThirtyEight’s archive, and the Internet Archive’s legal and technical struggles. It frames the issue as a cultural sovereignty problem and points to European efforts like Qwant and a German court ruling that held Google liable for AI-generated false statements as potential paths forward. Several claims in the essay rely on cited reporting rather than original research.

Filed from: The Walrus , Hacker News discussion .