Muse Glimmer 30B: Meta's distilled agent model runs on one consumer GPU
Distilling from a large teacher into a small model specialised for long tool-use chains may matter more than raw size for local agents, a trade-off the OSWorld gap makes visible. Builders running agents on local hardware can use it as a drop-in for multi-tool MCP workflows without cloud dependency. Compared with Qwen3.6-27B, it wins on tool orchestration but loses on general OS-level tasks.
Meta released a paper behind Muse Glimmer, a dense 30B model distilled from its larger Muse Spark and trained on synthetic agent runs in mock apps. Its 4-bit weights fit in under 20 GB, allowing it to run on a single RTX 5090 consumer GPU. On MCPAtlas, a multi-tool benchmark, it scored 75.5 versus 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B. It still trails Qwen3.6-27B on OSWorld-Verified, 65.9 against 75.6.
Meta's paper behind, Muse Glimmer 30B, a 30B model distilled from Meta's larger Muse Spark can run AI agents on 1 consumer GPU
Open models small enough for consumer hardware are usually general-purpose and weak at long tool-using tasks. That leaves local agents well behind cloud-hosted ones.
Meta distilled Muse Glimmer, a dense 30B model, from its larger Muse Spark and trained it on synthetic agent runs in mock apps. Its 4-bit weights fit in under 20 GB, so it runs on a single RTX 5090.
On MCPAtlas, a multi-tool test, it scored 75.5 against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B. It still trails Qwen3.6-27B on OSWorld-Verified, 65.9 against 75.6.
