Muse Glimmer 30B Model Card
by Meta
The reference card for Meta's Apache-2.0 30B agentic model, with the quantised memory footprints an engineer actually needs.
Overview
Muse Glimmer is Meta's open-weight agentic model, announced on 10 August 2026 and published under Apache 2.0 at huggingface.co/meta-models/Muse-Glimmer-30B, where the model card is the primary technical document. It pairs a 29.6B-parameter dense causal Transformer with a roughly 1.8B ViT-G/14 perception encoder for native image input, a context window of 131,072+ tokens, a 202,048-token vocabulary (200K BPE plus 2,048 special tokens) and a stated knowledge cutoff of 4 January 2026. The design target is an always-on local agent rather than a frontier chat model: a 4-bit build fits inside 24-32GB of consumer GPU memory, and the release ships a matched DFlash draft model for speculative decoding that Meta reports at roughly 3x speedup on an RTX 5090. Meta's accompanying research blog documents the training recipe in three phases — pre-training by logit distillation from the larger Muse Spark, mid-training on longer-context agent-heavy data, and post-training combining supervised fine-tuning with on-policy distillation and reinforcement learning. Reported results emphasise agentic rather than trivia benchmarks: MCP-Atlas 75.5%, SWE-Bench Pro 51.2%, CharXiv 78.8%, plus DeepSearch QA and tau-Bench, compared against Gemma4-31B and Qwen3.6-27B. The card documents multi-step workflow completion, schema-accurate function calling, recovery from tool failures, interleaved text-and-image input and coverage of 100+ languages, and notes integration with common agent orchestration frameworks. For a practising engineer the value is the specificity: it is one of the few current model cards that states quantised memory footprints, a bundled drafter model and agent-benchmark numbers in one place, which is exactly the combination you need to judge whether a local model can replace a hosted API inside an agent loop.
At a Glance
- Topic
- Models
- Level
- Intermediate
- Format
- Documentation
- Cost
- Free
- Duration
- ~30 min to read the card and the accompanying research blog; about a day to benchmark it on your own hardware
- Provider
- Meta
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How a 30B dense transformer plus a ViT-G/14 perception encoder is assembled for local agent work
- ✓What 4-bit quantisation of a 30B multimodal model actually costs in consumer GPU memory
- ✓How logit distillation from a larger teacher model differs from training a small model directly
- ✓Why mid-training on long-context agent-heavy data precedes SFT, on-policy distillation and reinforcement learning
- ✓How speculative decoding with a bundled drafter model changes local inference throughput in practice
- ✓Which benchmarks matter for an agentic model: MCP-Atlas, SWE-Bench Pro, tau-Bench and DeepSearch QA
- ✓How to read a model card critically for knowledge cutoff, vocabulary size, context length and licence terms
Highlights
- •Apache 2.0 rather than a bespoke community licence, a real change from the terms attached to earlier Meta model releases.
- •One of the few model cards that publishes quantised memory footprints alongside agentic benchmark scores, instead of leaderboard numbers alone.
- •Ships a matched DFlash draft model, so speculative decoding is a documented, supported path rather than a do-it-yourself project.
- •Simon Willison's independent hands-on writeup ran an 18.16 GB build on a 128 GB machine with room for other apps, and rated its image understanding notably strong while its drawing attempt came out jumbled.
- •609,919 downloads in its first month on Hugging Face, so community quantisations, bug reports and integration notes already exist.
- •Meta's research blog documents the full three-phase training recipe, a level of detail most open-weight releases skip entirely.
Who It’s For
Best For
- ✓Engineers evaluating whether a locally-run model can carry an agent loop without a frontier API
- ✓Teams with data-residency, latency or cost constraints that rule out hosted frontier models
- ✓Practitioners studying current open-weight post-training recipes that combine distillation with reinforcement learning
- ✓Anyone sizing consumer GPU hardware for on-device multimodal agents
Prerequisites
- •Familiarity with transformer terminology: dense versus mixture-of-experts, context length, vocabulary, KV cache
- •Practical experience running local models through llama.cpp, vLLM, Ollama or MLX
- •An understanding of quantisation formats and what 4-bit typically costs in output quality
FAQ
What is Muse Glimmer 30B Model Card?
The Hugging Face model card for Muse Glimmer 30B, Meta's open-weight model built for always-on local agent workflows and released under Apache 2.0 in August 2026. It is for engineers deciding whether a model they can run themselves is good enough to carry an agent loop. Read together with Meta's research blog it gives you the architecture, the context window, the quantised memory footprint, the speculative-decoding setup and the agentic benchmark scores needed to size hardware and set expectations before you download 30 billion parameters.
Is Muse Glimmer 30B Model Card free?
Muse Glimmer 30B Model Card is free to access.
What level is Muse Glimmer 30B Model Card for?
Muse Glimmer 30B Model Card is aimed at a intermediate audience. Recommended background: Familiarity with transformer terminology: dense versus mixture-of-experts, context length, vocabulary, KV cache, Practical experience running local models through llama.cpp, vLLM, Ollama or MLX, An understanding of quantisation formats and what 4-bit typically costs in output quality.
How long does Muse Glimmer 30B Model Card take?
Expect roughly ~30 min to read the card and the accompanying research blog; about a day to benchmark it on your own hardware. Most learners work through it at their own pace.
What will I learn from Muse Glimmer 30B Model Card?
You'll learn: How a 30B dense transformer plus a ViT-G/14 perception encoder is assembled for local agent work; What 4-bit quantisation of a 30B multimodal model actually costs in consumer GPU memory; How logit distillation from a larger teacher model differs from training a small model directly; Why mid-training on long-context agent-heavy data precedes SFT, on-policy distillation and reinforcement learning; How speculative decoding with a bundled drafter model changes local inference throughput in practice; Which benchmarks matter for an agentic model: MCP-Atlas, SWE-Bench Pro, tau-Bench and DeepSearch QA; How to read a model card critically for knowledge cutoff, vocabulary size, context length and licence terms.
Topics
Sources
This page was written from 3 sources, 2 on domains other than huggingface.co.