Meta Releases Muse Glimmer Bringing Local AI Agents to Consumer GPUs
Meta has launched the Muse Glimmer 30-billion-parameter model under an Apache 2.0 licence, targeting on-device workloads and private context processing.
Stock photo for illustration only, not from the actual event
- Meta has released the 30-billion-parameter Muse Glimmer model on Hugging Face under an Apache 2.0 licence.
- Designed to operate as local AI agents capable of running directly on consumer GPUs.
- Scores competitively against Gemma4-31B and Qwen3.6-27B across multiple general and coding benchmarks.
- Employs roughly 4-bit weight quantisation to shrink the language model under 20 GB for 24 GB or 32 GB memory envelopes.
Meta's Superintelligence Labs has published the weights for its new 30-billion-parameter model, Muse Glimmer, on Hugging Face under an Apache 2.0 licence. Developers can utilise the model for local coding, function calling, personal agents, and LLM-as-a-judge evaluation tasks.
This release addresses an operational bottleneck for AI teams, as cloud-hosted models require persistent network access and central infrastructure. Instead, Meta positions Muse Glimmer for workloads demanding on-device execution, including personal agents that interact directly with user schedules, messages, files, and other private contexts.
Stock photo for illustration only, not from the actual event
In general-agentic benchmark evaluations, Meta reports that Muse Glimmer outperforms Gemma4-31B and Qwen3.6-27B on five out of eight tests. The model achieved a score of 75.5 on MCP Atlas, compared to 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B. DeepSearch QA results show a similar trend with a score of 74.6, while the model also led SWE-Bench Pro evaluations in coding capabilities with 51.2 points.
Deploying a 30-billion-parameter model compressed below 20 GB marks a significant step toward practical on-device AI capabilities for consumer hardware. As modern desktop GPUs and high-end consumer systems increasingly feature 24 GB to 32 GB of VRAM, running sophisticated local agents becomes far more viable. However, connecting these autonomous systems to enterprise files or personal storage still requires rigorous permission boundaries and tool-use controls to mitigate unexpected execution errors.
Another key feature of Muse Glimmer is its dedicated perception encoder, which accepts interleaved text and images to allow agents to interpret screenshots, charts, and documents within a conversation. The model achieved a score of 78.8 on Charxiv Reasoning. Furthermore, Meta conducted safety-related evaluations including CI Memories and Siren AgentDojo to gauge violation rates and utility scores.
From a hardware perspective, a full-precision 30-billion-parameter model would normally demand over 55 GB of memory. Muse Glimmer utilises approximately 4-bit weight quantisation to reduce the language model footprint to under 20 GB, leaving sufficient headroom for a KV cache, perception encoder, and speculative-decoding drafter within a 24 GB or 32 GB memory envelope.
Source: AI News
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment