Meta drops 30B agentic model Muse Glimmer under Apache 2.0 — runs on a single GPU

Meta just open-sourced Muse Glimmer, a 30-billion-parameter model purpose-built for always-on local agent workflows. It’s distilled from Muse Spark and ships under Apache 2.0, which means solo developers, mid-market teams, and air-gapped enterprises all get the same day-one self-hosting path. No API key required.

The real engineering story here isn’t the parameter count — it’s the compression. A dense 30B transformer normally demands over 55 GB of memory at full precision. Meta squashes it to roughly 4-bit, bringing the language core under 20 GB. Two quantization builds are available: K-Quant-Dynamic for 32 GB VRAM setups with 0.2% average accuracy degradation, and K-Quant-17GB for 24 GB cards at 1.0% degradation. That’s measured across 15 benchmarks, so the tradeoff is quantified rather than hand-waved.

Speed comes from DFlash, a block-diffusion drafter that predicts 16 tokens in one forward pass while the main model verifies them in parallel. On an RTX 5090, throughput jumps from 74.9 to 233.4 tokens per second — a 3.1x improvement. Even Apple Silicon sees meaningful gains: the M5 Max climbs from 26.6 to 50.2 tok/s. These numbers matter because an agent loop dies if the model can’t respond fast enough. Meta is explicitly targeting desktop agents that read screenshots, coding assistants, and schema-based function calling. Healthcare, legal, financial services, defense, and manufacturing are the named verticals — any setting where data residency or latency kills a cloud dependency.

Benchmark positioning is clear-eyed. Muse Glimmer leads Gemma4-31B and Qwen3.6-27B on agentic orchestration and reasoning: 75.5 on MCP Atlas (versus 54.2 and 62.5), 74.6 on DeepSearch QA, and 94.7 on AIME 2026. But it trails Qwen on computer-use tasks — 65.9 versus 75.6 on OSWorld-Verified, and 60.7 versus 77.2 on TerminalBench 2.1. That split tells you exactly where this model fits. It’s a reasoning engine for structured agent pipelines, not a pixel-pushing operator. If your workflow chains tool calls, parses documents, and evaluates outputs, Glimmer is competitive. If you need a model to click buttons on a desktop, Qwen still has the edge. Meta also discloses a Siren AgentDojo attack success rate of 28.4 and rates chem/bio, cyber, and loss-of-control risk at moderate or lower. The company explicitly states the model doesn’t meet its internal Frontier AI threshold, which is either reassuring or limiting depending on your use case.