Muse Glimmer
Muse Glimmer is a 30-billion-parameter open-weight model from Meta Superintelligence Labs, released on August 10, 2026 under Apache 2.0 and built to run agents locally on a single GPU. Distilled from Muse Spark, it takes text and image input and returns text.
Released
August 10, 2026
Type
multimodal
License
open-weight
Apache 2.0
Context
—
Capabilities
Architecture
Parameters: 30B
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Meta AI Research: Introducing Muse Glimmer (opens in a new tab)
- Meta Model API docs: Muse Glimmer (opens in a new tab)
- Hugging Face: meta-models/Muse-Glimmer-30B model card (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: Muse Glimmer and Meta
MarkTechPost
Google Open-Sources TorchTPU, a Native PyTorch Backend for Cloud TPUs
Google has open-sourced TorchTPU, a native PyTorch backend for Cloud TPUs, co-developed with the PyTorch team at Meta. Announced at PyTorch Conference 2026 on October 20, 2026, TorchTPU enables existing PyTorch code to target a TPU through a standard torch.device("tpu"). TPUs are now recognized as a native PyTorch device alongside GPUs and CPUs. The Google team recommends TorchTPU for all new workloads, stating it replaces PyTorch/XLA. TorchTPU is a software stack comprising four open repositories: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan. Available as v0.2.0 on PyPI, it requires Python 3.11+ and is licensed under Apache 2.0. TorchTPU runs on Google Cloud TPUs and has been tested on generations from v5e, scaling up to 8,192 Ironwood (TPU7x) chips. The serving capabilities support FP8, MXFP4, and NVFP4 checkpoints. Performance metrics for TorchTPU include eager mode for debugging and torch.compile for production throughput. The system achieved a self-reported 52% MFU training for Qwen 3 8B with TorchTitan on Ironwood. In a fine-tuning scenario, out-of-the-box eager mode took 11.49 seconds per step, compared to 2.78 seconds when fully tuned. TorchTPU functions as an open-source PyTorch backend that allows PyTorch models to run natively on TPUs by plugging into PyTorch’s PrivateUse1 interface. Unlike its predecessor, PyTorch/XLA, which recorded lazy-tensor graphs and required developers to mark compilable regions, TorchTPU handles tensors as ordinary PyTorch tensors without subclasses or wrappers. It supports one process per device, preserves tensor views and in-place mutations, and handles divergent MPMD execution, whereas PyTorch/XLA only supported pure SPMD code. TorchTPU offers three eager modes: Strict (default), which dispatches each operator asynchronously with standard PyTorch semantics; Debug, which runs synchronously with tensor bounds checking; and DeferAndFuse, which fuses runs of eager operators to reduce HBM traffic. The TorchTPU paper indicates that DeferAndFuse can be 50% or more faster than Strict mode.
Search Engine Journal
What Personal AI Agents Like Meta Muse And OpenAI Dots Mean For SEO via @sejournal, @slobodanmanic
Personal AI agents like Meta's Muse and OpenAI's dots are accessing websites through connectors or cloud browsers. This integration is changing the effectiveness of robots.txt, on-page SEO, and analytics in understanding website traffic and content. The development signifies a shift in how search engines and AI tools interact with web content.
MarkTechPost
Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context
Meta Superintelligence Labs has released Muse Spark Open, the open-weight version of its Muse Spark model. This new model is an 816 billion parameter Mixture-of-Experts (MoE) model that activates 42 billion parameters per token. It is designed to process text, image, video, and audio inputs, returning text outputs, and boasts a 1 million token context window. Meta claims Muse Spark Open is the most capable US open-weight model released to date. For developers, a key feature is its deployment capability, with the NVFP4 build running on two Apple M5 Ultra Mac Studios. The model's architecture includes 62 layers and a hidden dimension of 8192, with each MoE layer containing 256 experts and activating 8 per token through top-k routing. The model supports multimodal input through a separate 1.8 billion parameter Perception Encoder. Its vocabulary size is 202,048 tokens, and it was trained on data covering over 100 languages. Muse Spark Open offers adjustable reasoning depth settings, defaulting to 'xhigh'. Performance benchmarks show Muse Spark Open excelling in agentic tool use and knowledge reliability, achieving a score of 89.2 on MCP-Atlas. However, it performed mid-pack on coding tasks, scoring 76.4 on Terminal-Bench, which is behind models like Kimi K3 (85.0), GLM 5.3 (83.9), and Qwen3.8 (82.0).
Hugging Face
Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
Meta has launched Muse Spark Open, a frontier multimodal mixture-of-experts AI model. The model is now available as open-weight, enabling wider access and use. This release signifies Meta's contribution to the advancement of open-weight AI models in the frontier technology space.
The Hacker News
Bring Your Own Agent: Your Employees' Assistants Are Already Logged In
The rollout of personal AI agents is encountering significant challenges, including privacy concerns and identification issues. Amazon has blocked Meta's Muse agent for failing to identify itself as a bot during browsing, and the Instinct agent has also raised privacy alarms. Despite these ongoing issues and accumulating negative experiences, the technology sector remains captivated by the capabilities and potential of these personal AI services. The article highlights that the personal AI agent rollout is proceeding with expected difficulties.