Meta's Muse Glimmer Runs a 30B Agent on One GPU

Meta released Muse Glimmer on Monday, a 30-billion-parameter open-weight model built specifically for agentic work that runs locally — and, unusually for a model of its class, it fits on a single consumer GPU. The weights went up on Hugging Face under an Apache 2.0 licence, which permits commercial use without a bespoke Meta agreement.

Architecturally, Glimmer is a distillation of the larger Muse Spark rather than a new pretraining run. It pairs a 2B ViT-style vision encoder with a 28B text decoder built on hybrid attention: 52 layers alternating 2,048-token sliding-window attention with rotary embeddings against full-attention layers using NoPE. Gated grouped-query attention shares each key-value head across 16 query heads, which is what cuts KV-cache memory by roughly 16x and makes the local deployment story viable. The context window is 32,768 tokens with an 8,192-token output ceiling. Meta describes a three-phase recipe in its research blog: logit distillation from the Spark teacher during pretraining, longer-context agent-heavy data in mid-training, then supervised fine-tuning combined with on-policy distillation and reinforcement learning.

The benchmark numbers Meta published put Glimmer ahead of the open models nearest its size. On MCP Atlas, which measures tool-calling against real Model Context Protocol servers, it scores 75.5 against 62.5 for Alibaba's Qwen3.6-27B and 54.2 for Google's Gemma4-31B. On SWE-Bench Pro it posts 51.2 against Qwen's 50.2 and Gemma's 36.9, and on AIME 2026 it reaches 94.7. The gap is widest precisely where the model is aimed — agentic tool use — and narrowest on the maths and coding benchmarks where the 27B-class field has already converged.

Memory is the part that matters for anyone planning to actually run it. Full weights need about 55 GB; calibrated 4-bit GGUF quantisation brings that under 20 GB, which fits a 24 GB or 32 GB card. Meta also ships an optional DFlash speculative-decoding drafter, and measured 3.1x faster generation on an Nvidia RTX 5090, 1.8x on an Apple M5 Max and 1.5x on an M4 Max, with the largest gains on structured output such as code. Day-zero support landed in transformers, llama.cpp and vLLM, with Ollama, LM Studio, ExecuTorch, MLX and SGLang listed as coming.

The release also reverses a direction Meta had appeared to be taking. After Llama 4 landed to a flat reception and the company stood up a new superintelligence organisation, its frontier work turned closed. Chinese labs filled the vacuum — Moonshot's Kimi K3, Alibaba's Qwen line and DeepSeek's V4-Flash now anchor the open-weight tier. Alongside Glimmer, Mark Zuckerberg pressed Washington to loosen constraints on training data and distillation, arguing that "foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," per TechStartups. Meta says weights for Muse Spark 1.2, its most capable model, will follow.

Whether Glimmer holds its lead depends on how the numbers survive contact with independent evaluation — Meta's own agentic benchmarks are the ones it optimised against, and MCP Atlas in particular rewards the tool-calling format a model was trained on. The more durable claim is the one about hardware: a 30B agent that quantises into 20 GB moves a whole category of always-on, privacy-bound workloads off the API and onto the desk.