Meta Launches Muse Glimmer, a 30B AI Model That Runs Locally—but Your Laptop May Not Be Ready
Meta has released Muse Glimmer, a 30-billion-parameter open-weight AI model designed to power agents directly on personal devices. It can handle tasks such as coding, calling software tools, interpreting images and completing multi-step workflows without depending continuously on cloud infrastructure.
However, Meta’s description of Muse Glimmer as a model for “consumer hardware” requires context: its practical configurations target high-end computers with approximately 24 GB or 32 GB of available memory, not the average everyday laptop.
Quick Answer
Muse Glimmer is Meta’s new open-weight model for running AI agents locally on a compatible Mac or PC.
Its weights are available under the permissive Apache 2.0 licence, allowing commercial and research use. Meta compressed the model to under 20 GB using approximately 4-bit quantization, but users still need additional memory for image processing, working context and speed-enhancing components.
The result is a model that can work without a constant internet connection—but only on sufficiently powerful hardware.
Key Takeaways
- Muse Glimmer contains approximately 30 billion parameters.
- It can process interleaved text and images but generates text-only output.
- The model is designed for agents, coding, function calling and AI evaluation.
- Its quantized configurations target systems with 24 GB or 32 GB of memory.
- Meta tested it on Apple M4 Max and M5 Max computers and an Nvidia RTX 5090.
- Meta reported speeds as high as 233.4 tokens per second on the RTX 5090 using speculative decoding.
- Benchmark and performance results were reported by Meta and have not necessarily been independently reproduced.
- The weights are available under the Apache 2.0 licence.
Built for more than answering questions
Muse Glimmer is designed for tasks that require more than producing a single chat response.
Meta trained and evaluated the model for end-to-end task completion, accurate tool use, multi-step reasoning and failure recovery. If a software tool fails or returns an unexpected result, the model is intended to diagnose the problem and attempt another approach instead of immediately stopping.
A dedicated perception encoder allows Muse Glimmer to accept text and images together. That means an agent could potentially interpret screenshots, charts and documents while following written instructions.
Meta identifies possible personal-agent uses such as managing schedules, drafting messages and organizing files. The model is compatible with OpenClaw and other agent-orchestration systems, offers selectable reasoning strengths and was trained on data covering more than 100 languages.
The model card lists a context length of at least 131,072 tokens, potentially allowing Muse Glimmer to work with lengthy conversations, documents or task histories. Its reported knowledge cutoff is January 4, 2026.
What running an AI model locally changes
Most advanced AI tools process requests through remote data centres. A local model can perform some work directly on a person’s computer, allowing it to operate without a constant internet connection and potentially reducing reliance on paid cloud-based AI services.
Local processing may also allow sensitive information to remain on the device instead of being sent to an external server. This could be valuable for agents that need access to personal files, messages, schedules or company documents.
However, local operation does not automatically make an AI agent secure. A model with broad access to files, email, calendars or software tools could still make mistakes, misunderstand instructions or encounter malicious content. Its permissions and actions would need to be carefully controlled.
“Consumer hardware” does not mean every laptop
A full-precision 30-billion-parameter model would require more than 55 GB of memory, according to Meta. That is beyond the capacity of most consumer graphics cards.
Meta used approximately 4-bit weight quantization to reduce the language model to under 20 GB. The company says this leaves enough room for the model’s KV cache, image-perception encoder and speculative-decoding drafter within a 24 GB or 32 GB memory envelope.
Meta offers two principal compressed configurations:
- K-Quant-17GB: Targets hardware with approximately 24 GB of memory.
- K-Quant-Dynamic: Targets hardware with approximately 32 GB of memory.
Meta reported average accuracy degradation of approximately 1% for the 17 GB version and 0.2% for the dynamic version across 15 benchmarks.
Those requirements place Muse Glimmer within reach of some individuals and developers, but largely through premium Apple Silicon computers, high-memory AI PCs or powerful desktop GPUs. A typical laptop with 8 GB or 16 GB of memory should not be assumed to deliver the experience Meta demonstrated.
How Meta made the model faster
Muse Glimmer includes a lightweight drafter model based on a technique called DFlash speculative decoding.
Ordinarily, a language model produces text one token at a time. The DFlash drafter instead proposes a block of tokens, which Muse Glimmer verifies in parallel. Correct proposals are accepted while incorrect ones are corrected by the main model.
Because the primary model still verifies the drafter’s suggestions, Meta says the technique accelerates generation without changing the output the main model would otherwise produce.
In Meta’s tests, the K-Quant-17GB model and quantized DFlash drafter delivered:
| Hardware | Without speculative decoding | With DFlash | Reported speed increase |
|---|---|---|---|
| Nvidia RTX 5090 | 74.9 tokens per second | 233.4 tokens per second | 3.1× |
| Apple M5 Max | 26.6 tokens per second | 50.2 tokens per second | 1.8× |
| Apple M4 Max | 23.7 tokens per second | 37.8 tokens per second | 1.5× |
Meta conducted these tests using a batch size of one and greedy decoding. Apple testing used ExecuTorch, while the RTX 5090 test used llama.cpp. Actual performance may differ depending on the computer, configuration, prompt and software implementation.
How Muse Glimmer performed in benchmarks
Meta compared Muse Glimmer with other similarly sized open-weight models, including Gemma4-31B and Qwen3.6-27B.
In Meta’s testing, Muse Glimmer outperformed both competing models on several agentic evaluations, including MCP Atlas, DeepSearch QA and WildClawBench. However, it did not lead every benchmark. Qwen3.6-27B scored higher on tests including GDPVal-AA v2, SkillsBench and OSWorld-Verified.
This makes the results more nuanced than saying Muse Glimmer is simply the strongest model in its class. Its apparent advantage is concentrated in certain tool-using and agentic tasks, while competing models remain stronger in other areas.
The comparisons should also be treated as company-reported results. Meta used internal reproductions for some third-party-model evaluations, and its methodology notes that those testing environments might not be optimized for every competing model.
Where developers can access Muse Glimmer
Meta has released the Muse Glimmer weights through Hugging Face under the Apache 2.0 licence.
Optimized integrations with llama.cpp, MLX and ExecuTorch are expected in the coming days. Meta also identifies Ollama, LM Studio and Unsloth as options for running the model locally, while vLLM and SGLang can be used for serving it at scale.
The company is working with AMD, Arm, Dell, Intel and Nvidia to optimize performance across additional devices.
TwikUp Insight
Muse Glimmer’s most important feature may not be that it beats every similarly sized AI model—it does not. The more consequential development is that an agent capable of reasoning, interpreting images, calling tools and recovering from failures can now fit within the memory limits of a powerful personal computer.
That could shift certain AI workflows away from cloud servers and onto users’ own devices, reducing recurring cloud costs and giving people greater control over where their information is processed.
But the phrase “runs on your device” should not be confused with “runs on every device.” For now, Meta’s vision of an always-available local AI agent still requires hardware that many consumers do not own.