Muse Glimmer 30B: Features, Benchmarks & Local Setup

In short: Muse Glimmer 30B is a dense language model from Meta with about 29.6 billion parameters, open weights under Apache 2.0, and a 131,072-token context. It takes text and images and is built for local agent workloads. In Meta’s own table, it wins 12 of 22 comparisons against Gemma 4 31B and Qwen 3.6 27B. Artificial Analysis’s independent measurement places it behind Qwen, with an 82% hallucination rate. At 4-bit, it needs about 17 GB of memory, and third-party providers charge roughly $0.30 per million input tokens. We did not benchmark the model ourselves, so we keep vendor claims and independent numbers apart.

What is Muse Glimmer 30B?

Muse Glimmer 30B is Meta’s first open-weight model since Llama 4. It appeared on Hugging Face on 10 August 2026 and is the first open release from the restructured Superintelligence Labs. According to Meta, it was distilled from the larger proprietary Muse Spark. It belongs to a wider family, covered in our posts on Muse Code and Muse Spark 1.2 and on Meta’s strategic shift with Muse Spark. This guide focuses on the small model you can run locally.

Specs

Is Muse Glimmer 30B open source?

The accurate term is open weights under Apache 2.0. Unlike Llama, there is no monthly-active-user cap. A usage policy applies on top and excludes, for example, weapons, fraud, and use by anyone under 18. Training data and the weights of the teacher model Muse Spark are not released, so “fully open source” overstates it. Why large vendors back open models is covered in our post on Meta, Microsoft and Nvidia.

Muse Glimmer 30B benchmarks

Two things need to stay apart: vendor claims and independent measurement.

Vendor numbers (Meta)

Meta compares against Gemma 4 31B and Qwen 3.6 27B, both in thinking mode, and picks the comparison models itself. Muse Glimmer wins 12 of 22 rows and loses 10.

The strength is tool-using work: MCP Atlas, τ³-Banking and SWE-Bench Pro. On terminal and computer control, Qwen is clearly ahead.

Independent measurement (Artificial Analysis)

At launch, Muse Glimmer ranked 5th of 140 open models in the 4B to 40B class. The scale was later recalibrated. Per OrcaRouter, Qwen 3.6 27B then scores 21, Muse Glimmer 17, and Gemma 4 31B 15, with the same order.

Context: Only the Intelligence Index is fully independent. The agent results are Meta’s own. Meta itself notes that the model can make errors in multi-step reasoning.

Muse Glimmer 30B vs Qwen 3.8 27B

Most published numbers are against Qwen 3.6 27B. Artificial Analysis now lists that model as replaced by Qwen 3.8 27B. We found no direct measurements of Muse Glimmer against Qwen 3.8 27B, although many people search for them. We will add the comparison once reliable data exists. For the larger Qwen model, read Qwen 3.8 Max and Alibaba Qwen 3.

Against Qwen 3.6 27B:

• Muse Glimmer ahead: tool use (MCP Atlas), banking agent (τ³), SWE-Bench Pro.

• Qwen ahead: Terminal-Bench, OSWorld, GDPval-AA, hallucination rate, and the overall Intelligence Index.

Muse Glimmer 30B pricing

Meta runs no endpoint of its own for this model. Cost depends on the provider.

Prices change quickly. Check the provider’s current price page before you commit.

How to run Muse Glimmer 30B locally

Hardware requirements

Memory should exceed the file size. Otherwise llama.cpp offloads to RAM or disk and slows down sharply. Sources put the BF16 size between 55.7 and 64 GB.

Run it with llama.cpp

Unsloth recommends temperature 1.0, top_p 0.95, and top_k 64.

For text-only mode:

bash

./llama.cpp/llama-cli \

    -hf unsloth/Muse-Glimmer-30B-GGUF Muse-Glimmer-30B-UD-Q4_K_XL.gguf \

    --temp 1.0 --top-p 0.95 --top-k 64

For running the API Server variant (llama-server):

bash

./llama.cpp/llama-server \

    -hf unsloth/Muse-Glimmer-30B-GGUF Muse-Glimmer-30B-UD-Q4_K_XL.gguf \

    --mmproj mmproj-BF16.gguf \

    --host 127.0.0.1 --port 8080

For images, also load the vision file (--mmproj mmproj-BF16.gguf). For an API server, use llama-server with the same model.

Ollama and LM Studio

The Ollama library has an entry muse-glimmer:30b, so the command is ollama run muse-glimmer:30b. Check the tag and size before downloading. For how the two tools differ, see Ollama vs LM Studio. GGUF files can generally be loaded in LM Studio too, but we found no official guide for this model.

Speed

With the DFlash draft model (speculative decoding), Meta cites about a 3.1x speedup on an RTX 5090, from 74.9 to 233.4 tokens per second. On an M5 Max, it is 1.8x; on an M4 Max, 1.5x. DataCamp’s own llama.cpp test on an RTX 5090 with 64K context measured about 84 to 121 tokens per second at roughly 24 GB of VRAM. The setups differ, so do not compare the numbers directly. More practical examples are in lightweight LLMs on a single GPU.

What’s new

• Muse Spark Open: Meta released a much larger open model (816B parameters, 42B active, 1M context). It uses its own Muse license rather than Apache 2.0.

• Technical report: Muse Glimmer was trained by logit distillation from Muse Spark. The 4-bit variant averages 74.8% across 15 benchmarks, and the BF16 model 75.7%.

Both points come from a single source that sells competing access, and neither is independently confirmed.

Limits and risks

• Hallucinations: 82% in the Artificial Analysis measurement, versus 49% for Qwen 3.6 27B.

• Terminal and computer control: behind Qwen.

• Debugging: DataCamp calls the model fast at building but weak at diagnosing failures. Vague tasks worked poorly; numbered steps worked better.

• Security: In the AgentDojo test, 28.4% of prompt-injection attempts succeeded (Gemma 4: 25.6%, Qwen 3.6: 40.3%, Meta figures). That is not a security guarantee.

• Vendor numbers: Meta picks the comparison models and settings.

Who should use Muse Glimmer 30B?

Yes for local agents with tool calls, coding assistants on a 24 to 32 GB GPU, teams that do not want to send data to cloud services, and anyone who needs a permissive license. For a coding-agent example, see Kilo Code.

Probably not for tasks that demand high accuracy where hallucinations are costly, or for terminal and desktop automation, where Qwen leads. Other options are Kimi K3 and DeepSeek V4. Meta’s road to open models began with LlamaCon 2025 and the Llama 4 delay.

Bottom line

Muse Glimmer 30B is a strong open agent model for hardware with 24 to 32 GB of memory, but not a clear winner against Qwen. It wins on tool use, loses on terminal tasks, and its hallucination rate is high. If you want to build local agents, try it. If you need maximum reliability, compare it with Qwen 3.8 27B.

FAQ

What is Muse Glimmer 30B?
A dense language model from Meta with about 29.6 billion parameters, open weights (Apache 2.0), and image understanding.

When was it released?
On 10 August 2026 on Hugging Face. A technical report followed later.

How much RAM or VRAM does it need?
About 17 GB at 4-bit, 20 to 22 GB at 6-bit, and roughly 56 to 64 GB at full BF16, depending on the source.

Is it better than Qwen 3.6 27B?
It depends on the task. Muse Glimmer leads on tool use; Qwen leads on terminal, computer control, and hallucination rate.

Is it free?
The weights are open. Meta offers no API of its own. Third parties charge about $0.30 per million input tokens.

What is the context size?
131,072 tokens, up to 262,144 per Unsloth.

Sources

Artificial Analysis

Qubrid

Unsloth Docs

OrcaRouter

Mer.vin

DataCamp