Meta Superintelligence Lab writes that Muse Glimmer-30B is a 30-billion-parameter vision-language model optimized for autonomous agentic workflows on consumer-grade hardware. The architecture combines a dense causal transformer with a dedicated ~1.8-billion-parameter vision encoder to process interleaved text and images, enabling multi-step planning, reliable tool invocation, and automatic error recovery. Designed to run locally without cloud dependency, the model employs 4-bit quantization and a novel DFlash speculative decoding drafter to achieve significant speedups on devices with 24 to 32 GB of VRAM. Evaluated against comparable 27 to 31 billion parameter systems, Muse Glimmer demonstrates strong performance across agentic, coding, and multimodal reasoning benchmarks while maintaining strict safety guardrails and supporting over 100 languages.
- Trained on data curated from public sources, third parties, and Meta's internal products, with a knowledge cutoff of January 2026.
- Supports controllable reasoning strength (low, medium, high, xhigh) to balance output quality and inference speed.
- Includes a frozen ViT-G/14 perception encoder and releases both full-precision BF16 weights and two 4-bit quantized variants.
- Recommended inference settings include a temperature of 1.0, top-p of 0.95, and top-k of 64.
- Assessed for moderate or lower risk in cyber, loss-of-control, and chemical/biological domains, though explicit safety guardrails are still recommended for deployment.
- Trained on data curated from public sources, third parties, and Meta's internal products, with a knowledge cutoff of January 2026.
- Supports controllable reasoning strength (low, medium, high, xhigh) to balance output quality and inference speed.
- Includes a frozen ViT-G/14 perception encoder and releases both full-precision BF16 weights and two 4-bit quantized variants.
- Recommended inference settings include a temperature of 1.0, top-p of 0.95, and top-k of 64.
- Assessed for moderate or lower risk in cyber, loss-of-control, and chemical/biological domains, though explicit safety guardrails are still recommended for deployment.
Pedro Cuenca writes Meta released Muse Glimmer-30B, a local, open-source multimodal model distilled from its larger Muse architecture. Designed for agentic workflows, it combines a 28B text decoder with a 2B vision encoder, supporting image, video, and multimodal tool calling out of the box. The release includes immediate compatibility with major inference frameworks like transformers, llama.cpp, and vLLM, alongside built-in speculative decoding for faster generation.
- Features a hybrid attention pattern alternating between three sliding window layers and one full attention layer.
- Incorporates a DFlash block-diffusion drafter to accelerate structured text generation like coding.
- Supports fine-tuning via TRL with practical minimums ranging from one to eight H100 GPUs depending on the method.
- Demonstrates autonomous agent capabilities such as self-quantization, self-deployment, and hardware-specific optimization.
- Features a hybrid attention pattern alternating between three sliding window layers and one full attention layer.
- Incorporates a DFlash block-diffusion drafter to accelerate structured text generation like coding.
- Supports fine-tuning via TRL with practical minimums ranging from one to eight H100 GPUs depending on the method.
- Demonstrates autonomous agent capabilities such as self-quantization, self-deployment, and hardware-specific optimization.
Meta is addressing high DDR5 memory costs by repurposing legacy DDR4 modules from decommissioned servers. Through a custom-developed Vistara ASIC, the company can attach old DDR4 memory to modern servers running AMD EPYC Turin processors that natively only support DDR5. This CXL 2.0 implementation allows for expanded memory capacity by using DDR4 as a secondary, slower tier for cold data while keeping frequent data in fast DDR5.
- Meta's Vistara ASIC bridges legacy DDR4 with modern DDR5 servers
- CXL 2.0 technology enables tiered memory management via NUMA nodes
- Panmnesia offers scalable CXL controller and switch solutions for data centers
- Strategy aims to mitigate rising DRAM prices and hardware costs
- Meta's Vistara ASIC bridges legacy DDR4 with modern DDR5 servers
- CXL 2.0 technology enables tiered memory management via NUMA nodes
- Panmnesia offers scalable CXL controller and switch solutions for data centers
- Strategy aims to mitigate rising DRAM prices and hardware costs
This article explores how Meta's leadership has shifted its engineering culture toward an intense focus on large language model development, often at the expense of core stability and employee autonomy. The author details a series of controversial management decisions, including the forced reassignment of thousands of engineers to manual data labeling and reinforcement learning from human feedback tasks. These changes have allegedly led to invasive employee monitoring through keystroke tracking and created incentives for "tokenmaxxing," where developers use generative tools excessively just to boost performance metrics. Such organizational shifts are also linked to significant security lapses, including major Instagram account takeover incidents caused by reliance on automated code reviews and understaffed security teams.
* Shift from engineering autonomy toward LLM-centricity
* Reassignment of engineers to repetitive data labeling and RLHF work
* Implementation of invasive keystroke and mouse tracking for training data collection
* Performance metric distortion through excessive token usage inflation
* Correlation between security team downsizing and major service outages
* Shift from engineering autonomy toward LLM-centricity
* Reassignment of engineers to repetitive data labeling and RLHF work
* Implementation of invasive keystroke and mouse tracking for training data collection
* Performance metric distortion through excessive token usage inflation
* Correlation between security team downsizing and major service outages
Meta’s new “semi-formal reasoning” technique boosts LLM accuracy for code tasks (review, bug detection, patching) by having the AI reason through code instead of running it. This involves stating assumptions, tracing steps, and drawing conclusions – a structured process that improves results (up to 93% accuracy) and lowers computing costs.
Meta is heavily investing in AI integration, demonstrated through "AI Week" – intensive training sessions for employees. These weeks involve hackathons, demos, and hands-on experimentation with tools like Anthropic's Claude Code. The goal is to foster AI adoption across all job functions and seniority levels, with a focus on AI agents capable of automating tasks like coding and report generation.
Meta is also restructuring teams into AI-native "pods" and setting specific AI adoption targets. CEO Mark Zuckerberg believes 2026 will see a significant impact of AI on the way Meta employees work, despite recent layoffs and the delayed launch of its own AI model.
Meta is also restructuring teams into AI-native "pods" and setting specific AI adoption targets. CEO Mark Zuckerberg believes 2026 will see a significant impact of AI on the way Meta employees work, despite recent layoffs and the delayed launch of its own AI model.
Cisco and Meta are championing open-source large language models (LLMs) for enterprise threat defense, announcing new models and initiatives at RSAC 2025. Cisco's Foundation-sec-8B LLM and Meta's AI Defenders Suite aim to provide scalable, secure, and cost-effective cybersecurity solutions through collaboration and open innovation.
Newsweek interview with Yann LeCun, Meta's chief AI scientist, detailing his skepticism of current LLMs and his focus on Joint Embedding Predictive Architecture (JEPA) as the future of AI, emphasizing world modeling and planning capabilities.
The article discusses the release of Llama 3.2, a new model from Meta, and explores its capabilities and limitations, particularly focusing on its availability and usage for personal projects. The article is a light-hearted take on exploring new AI technologies for creative and personal endeavors.
Meta AI has released quantized versions of the Llama 3.2 models (1B and 3B), which improve inference speed by up to 2-4x and reduce model size by 56%, making advanced AI technology more accessible to a wider range of users.
