AI Tools

What Is Muse Glimmer and Can You Run It Locally?

Published 2026-08-11 · Researched and price-checked on official sources

Muse Glimmer is Meta's 30-billion-parameter open agentic model, released August 10, 2026 under an Apache 2.0 license. Its 4-bit weights fit under 20 GB, so it runs on a single consumer GPU with 24 GB of VRAM, and OpenRouter lists it at $0.35 per million input tokens. It targets tool calling, long-horizon tasks, and image understanding.

Key facts

What matters

Muse Glimmer is a 30-billion-parameter open agentic model from Meta Superintelligence Labs, Apache 2.0 licensed and released August 10, 2026.
4-bit quantized weights take under 20 GB: Meta targets 24 GB and 32 GB consumer GPUs, while full precision needs 64 GB of VRAM.
Context window is 131,072 tokens with text and image input, and the knowledge cutoff is January 4, 2026.
Free to run locally, or $0.35 per million input and $1.50 per million output tokens on OpenRouter; Meta's cloud model Muse Spark costs $1.25 and $4.25 per million.
On Meta's benchmarks, Muse Glimmer beats Gemma4-31B on MCP Atlas (75.5 vs 54.2) and SWE-Bench Pro (51.2 vs 36.9), while Qwen3.6-27B edges it on SWE-Bench Verified (77.2 vs 76.0).

What is Muse Glimmer?

Muse Glimmer is a 30-billion-parameter open agentic model from Meta Superintelligence Labs, distilled from the larger Muse Spark model and released as open weights under an Apache 2.0 license on August 10, 2026. Meta positions it as an always-on local agent model: it chains multi-step reasoning, calls tools with precise schemas, recovers from failed tool calls, and accepts images alongside text through a dedicated perception encoder.

Meta's model card lists local agents, coding agents, tool use, multimodal reasoning, synthetic data generation, and LLM-as-a-judge evaluation as intended uses, and states the model works with agentic orchestration tools like OpenClaw. It is trained on data from more than 100 languages.

Muse Glimmer specs: what you get

Parameters: roughly 29.6 billion, including a 1.8-billion-parameter ViT-G/14 perception encoder for images.
Context length: 131,072 tokens, using a hybrid local-global attention pattern with a 2,048-token sliding window.
Modalities: text and image input, text output. Audio is not supported.
Vocabulary: 202,048 tokens (200,000 BPE plus 2,048 special tokens).
Knowledge cutoff: January 4, 2026.
License: Apache 2.0, covering the full-precision weights, two 4-bit quantized variants, the DFlash speculative-decoding drafter, and the perception encoder.

Hardware requirements: how much GPU do you need?

At full precision a 30-billion-parameter model needs over 55 GB of memory, which no consumer GPU offers. Meta ships two 4-bit quantized variants that shrink the language model to under 20 GB, leaving room for the KV cache, the perception encoder, and the speculative decoding drafter.

Full precision (BF16): 64 GB VRAM, the baseline for the degradation figures.
K-Quant-Dynamic: 32 GB VRAM, 0.2% average degradation across 15 common benchmarks.
K-Quant-17GB: 24 GB VRAM, 1.0% average degradation.

Meta measured the K-Quant-17GB model with the DFlash drafter at 233.4 tokens per second on an Nvidia RTX 5090 (3.1x faster than no speculation), 50.2 tok/s on an Apple M5 Max, and 37.8 tok/s on an M4 Max, using batch size 1 and greedy decoding.

Muse Glimmer vs Gemma4-31B vs Qwen3.6-27B: benchmarks

The numbers below are Meta's own results, published with the model card on August 10, 2026. Muse Glimmer leads its size class on general agentic tasks like MCP Atlas and DeepSearch QA, while Qwen3.6-27B wins several coding and screen tasks. All three are open weight: Gemma4-31B-it and Qwen3.6-27B are both listed with the Apache 2.0 license on Hugging Face, matching Muse Glimmer.

Muse Glimmer pricing: free weights, or $0.35 per million tokens

The weights are free. Apache 2.0 lets you run, modify, and redistribute Muse Glimmer, including commercially. What you pay is hardware if you run it locally, or per token if you use a hosted provider.

OpenRouter lists Muse Glimmer at $0.35 per million input tokens and $1.50 per million output tokens, with cached input at $0.04 per million and a 131,072-token context. For comparison, Meta's larger cloud model Muse Spark 1.2 costs $1.25 per million input and $4.25 per million output tokens on OpenRouter, with a 1,048,576-token context window. Together AI and Fireworks AI are also named as launch partners by Meta; their per-token prices were not verified for this article.

Limitations and the honest catch

Muse Glimmer is a 30B model, not a frontier model. Meta states it does not fall under its own definition of Frontier AI, being generally less capable than Muse Spark, and rates its chem/bio, cyber, and loss-of-control risk moderate or lower. For agentic use, Meta recommends additional guardrails such as human-in-the-loop confirmation for irreversible actions.

Video input is processed as individual frames; audio input and output are not supported.
The knowledge cutoff is January 4, 2026, and performance degrades on languages outside the strongly supported set.
Quantized inference can show minor quality differences in edge cases versus full precision.
llama.cpp, MLX, and ExecuTorch integrations were announced as landing in the days after launch, so day-one local setup may still be settling.
Meta says the model is not intended for use by anyone under 18.

Who should use Muse Glimmer?

Use it if you want agent capabilities without per-token API bills: a 24 GB or 32 GB GPU turns Muse Glimmer into a private coding and tool-use agent that works offline. Developers building products that must keep user data on-device get a permissively licensed model they can fine-tune and redistribute.

Skip it if you need frontier reasoning for complex planning, a 1M-token context window, or plug-and-play cloud reliability. Those cases point back to paid API models, including Meta's own Muse Spark at $1.25 per million input tokens.

At a glance

Benchmark	Muse Glimmer 30B	Gemma4-31B	Qwen3.6-27B
MCP Atlas (public)	75.5	54.2	62.5
DeepSearch QA	74.6	61.7	71.1
SWE-Bench Pro	51.2	36.9	50.2
SWE-Bench Verified	76.0	66.6	77.2
Tau3-Banking	23.5	15.1	16.7
Gaia2	43.3	36.4	40.0
AIME 2026	94.7	89.2	94.1
IFBench	77.0	76.0	70.8

Muse Glimmer 30B vs Gemma4-31B vs Qwen3.6-27B, from Meta's model card, August 10, 2026 (higher is better)

FAQ

Is Muse Glimmer free?

Yes. Meta released the weights under Apache 2.0, so Muse Glimmer is free to download, run, modify, and use commercially. Running it locally requires a GPU with 24 GB or more of VRAM; hosted use starts at $0.35 per million input tokens on OpenRouter.

What GPU do I need to run Muse Glimmer?

Meta's 4-bit K-Quant-17GB variant fits in 24 GB of VRAM with 1.0% average benchmark degradation, and K-Quant-Dynamic targets 32 GB with 0.2%. Full precision needs 64 GB. On an RTX 5090, Meta measured 233.4 tokens per second with the DFlash drafter.

How does Muse Glimmer compare with Muse Spark?

Muse Spark is Meta's larger cloud model and the teacher Muse Glimmer was distilled from. On OpenRouter, Muse Spark 1.2 costs $1.25 per million input and $4.25 per million output tokens with a 1M-token context; Muse Glimmer costs $0.35 and $1.50 per million and runs locally on consumer GPUs.