Meta Releases Muse Glimmer, a 30B Multimodal Model Distilled for Local Agentic Use

Hugging Face's Transformers library shipped version 5.15.0 today, headlined by the addition of Meta Muse Glimmer, a new 30-billion-parameter multimodal model built for agentic use cases that can be run locally.

Muse Glimmer: distilled for local deployment

According to Meta, Muse Glimmer is distilled from a larger model called Muse down to 30B total parameters, split into two components:

• A 2B parameter ViT-style vision encoder (Perception Encoder)
• A 28B parameter text decoder

The model is released under the Apache 2.0 license, making it freely available for commercial use and local deployment. Meta positions it for privacy-sensitive applications that need to run on local hardware rather than through cloud APIs — including coding assistants, document analysis, and personal-assistant setups comparable to existing Claude-like or Hermes-like agent stacks.

Meta has not disclosed a context window size, pricing (irrelevant for a self-hosted Apache 2.0 release), or standardized benchmark scores in the release notes. Technical details are covered in a companion blog post published by Hugging Face at hf.co/blog/muse-glimmer, though full evaluation numbers were not included in the GitHub release itself.

Other model additions in v5.15.0

The release also adds support for two other model families:

GraniteSWA and GraniteMoeSWA — IBM-originated architectures (Granite with sliding window attention, in both dense and mixture-of-experts variants) gained native Transformers support via pull request #47179, contributed by developer @daviswer. Documentation pages for both granite_swa and granitemoe_swa are now live on the Hugging Face docs site.

A.X-K1 and A.X-K2 — additional model architectures were added to the library, though the release notes provided in this update do not include parameter counts, licensing terms, or benchmark data for these two models.

What this means

Muse Glimmer is the notable release here: a 30B multimodal model small enough to run on a single high-end GPU or a well-specced workstation, distilled from a larger Meta model and released permissively under Apache 2.0. That combination — local deployability, vision+text capability, and an open license — targets the same niche occupied by models like Llama's smaller variants and open Hermes fine-tunes, but aimed specifically at agentic workloads (coding, document processing, assistants) rather than general chat.

The lack of published benchmark scores or a stated context window is a real gap. Until Meta or independent evaluators publish standardized numbers (MMLU, GPQA, coding benchmarks, or vision-language benchmarks like MMMU), claims about Muse Glimmer's capability relative to other 30B-class open models — such as Qwen or Mistral's mid-size releases — can't be verified. The broader v5.15.0 release also reflects the steady expansion of Transformers as the default integration point for new open model architectures, with IBM's Granite-SWA variants and the A.X-K series adding to an already crowded field of open-weight releases arriving through the library each month.