Meta Publishes Technical Report for Muse Glimmer 30B, Its On-Device Agent Model
Meta posted the technical report for Muse Glimmer 30B, its 30-billion-parameter on-device agent model, revealing the training and quantization details behind the Apache 2.0 weights released in August.
By Icarus3 min read
Meta Publishes Technical Report for Muse Glimmer 30B, Its On-Device Agent Model
The paper gives the full technical record for Meta’s on-device agent model — but the model itself isn’t new, so treat this as documentation, not a launch.
Meta has posted the technical report for Muse Glimmer 30B, its 30-billion-parameter model built for autonomous agentic tasks on consumer hardware. The paper, titled “Muse Glimmer 30B: An On-Device Agentic Model,” appeared on Meta’s research publications page on October 19 and was submitted to arXiv on October 17 as 2610.24628.
This is a paper posting, not a launch. The Glimmer 30B weights themselves were already open-sourced under the Apache 2.0 license in August, and Meta has pointed developers at a Hugging Face download and documentation since then. What’s new is the written technical record: how Meta Superintelligence Labs built and trained the model, and the engineering choices that let a 30B model run locally instead of in a data center.
The Model, On Paper
The report describes an open-weight 30B-parameter causal language model with a perception encoder, built for autonomous agentic tasks on consumer hardware. The model combines multi-step reasoning, tool use, and failure recovery, takes text and image inputs, and produces text. It was trained on data from over 100 languages, and the report positions it for always-on local agent workflows: function calling, local coding, and LLM-as-a-judge evaluation.
The memory budget is the core engineering claim. At full precision, a 30B model needs over 55 GB of memory — more than any consumer GPU offers. Meta quantizes the model to roughly 4-bit precision, shrinking the language model to under 20 GB, leaving room for the KV cache, the perception encoder, and a speculative-decoding drafter inside a 24 GB or 32 GB envelope. The report claims minimal to no degradation on agentic tasks from the compression. A DFlash-based drafter proposes blocks of tokens for the main model to verify; Meta reports decode speedups of 3.1x on an RTX 5090, 1.8x on an M5 Max, and 1.5x on an M4 Max on a 17 GB quantized build.
The report says pre-training used the outputs of Muse Spark with logit distillation and a similar data mix to the teacher. Mid-training used longer-context, more agent-heavy data with richer reasoning traces, plus organic data. Post-training combined supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains. Meta says the model was evaluated under its Advanced AI Scaling Framework and assessed for open-weight release.
Benchmarks, Labeled as Meta’s
Meta claims the model is state of the art for its size class on tool use, multi-step planning, and agentic task completion, versus Gemma4-31B and Qwen3.6-27B. Meta’s reported figures have Glimmer 30B at 75.5 on the MCP-Atlas agentic benchmark, against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B, and at 74.6 on DeepSearch QA. Independent checks of the released weights have found the lead narrower than the headline claim: public leaderboard compilations show Qwen3.6-27B ahead on computer-use and terminal-based tasks, including OSWorld-Verified and TerminalBench 2.1, while Gemma4-31B leads on GPQA Diamond. Those comparisons are Meta’s on one side and community evaluations on the other; treat both as provisional.
Why It Matters
Glimmer 30B is part of Meta’s on-device strategy for Muse: agent models small enough to run on a single consumer GPU or a Mac instead of a cloud data center. That’s the counterpart to the cloud-scale releases — Muse Wonder, Meta’s frontier model family, and Muse Spark Open, the 816B open-weight model tuned for two Mac Studios. The technical report turns Glimmer from a downloadable artifact into a documented one, which matters for developers and researchers weighing whether to build local agent workflows on top of it. No Meta-hosted API exists; developers self-host or use third-party providers such as Together AI and Fireworks AI.
