Meta Platforms (Nasdaq: META) has introduced Muse Glimmer, a 30-billion-parameter artificial intelligence model designed to power always-on agentic workflows directly on consumer hardware.
Developed by Meta Superintelligence Labs, Muse Glimmer is available with open model weights under the permissive Apache 2.0 license. Meta said the model is compact enough to run locally on a Mac or PC equipped with a single compatible consumer graphics processor.
The release targets applications including local AI agents, function calling, software development and “LLM-as-a-judge” evaluations. It also reflects Meta’s broader effort to reduce developers’ dependence on cloud-based AI infrastructure.
The model weights and technical information are available through Meta’s official Muse Glimmer repository on Hugging Face.
What Is Meta Muse Glimmer?
Muse Glimmer is a dense, multimodal model with approximately 29.6 billion parameters, including its vision encoder.
The model is designed to process interleaved text and images, allowing applications to interpret screenshots, charts and documents alongside written instructions.
According to Meta’s official announcement, Muse Glimmer has been optimized for several agent-related capabilities:
- End-to-end completion of complex tasks
- Reliable function calling and tool use
- Multi-step reasoning across extended workflows
- Recovery from failed or unexpected tool responses
- Coding, debugging and software-development tasks
- Analysis of text and visual inputs
- Adjustable reasoning effort
- Support for more than 100 languages
Meta said the model can work with OpenClaw and other agent orchestration frameworks.
Muse Glimmer Runs Locally on Consumer Hardware
One of Muse Glimmer’s main selling points is its ability to operate without depending continuously on cloud infrastructure or an internet connection.
Running an AI model locally can provide several potential benefits, including reduced latency, greater control over sensitive information and lower recurring cloud-computing costs. Actual performance and privacy, however, will depend on the hardware, software configuration and applications connected to the model.
At full precision, Meta estimates that a model of this size would require more than 55 GB of memory. The company uses quantization to compress Muse Glimmer’s language-model weights to approximately four-bit precision, reducing the model to less than 20 GB.
This allows the quantized model, its working memory, perception encoder and speculative-decoding system to operate within an estimated 24 GB or 32 GB memory envelope.
Advanced Micro Devices (Nasdaq: AMD) said Muse Glimmer can run on selected Ryzen AI Max+ systems and Radeon AI Pro graphics cards. In preliminary company testing, the model reached up to 24 tokens per second on a Ryzen AI Max+ 395 system and up to 53 tokens per second on a Radeon AI Pro R9700 with DFlash enabled. These results are hardware- and configuration-specific and may not represent performance in other environments. AMD published its testing details and system configurations.
Meta Uses DFlash to Accelerate AI Generation
Muse Glimmer includes a lightweight speculative-decoding model based on DFlash.
Traditional language models generally generate responses one token at a time. Speculative decoding attempts to accelerate this process by allowing a smaller companion model to propose blocks of tokens.
The main model reviews those proposed tokens in parallel, accepting valid predictions and correcting inaccurate ones. Meta says this approach increases generation speed without changing the intended output quality.
Quantized versions of the companion model are also included to limit additional memory requirements.
How Meta Trained Muse Glimmer
Meta trained Muse Glimmer using knowledge transferred from Muse Spark, a larger teacher model.
The training process involved three principal stages:
Pre-Training
Meta trained Muse Glimmer on outputs generated by Muse Spark using logit distillation. This process allows a smaller model to learn from the probability distributions produced by a larger model.
Mid-Training
The company introduced longer-context and more agent-focused training data, including richer reasoning traces and organic data.
Post-Training
Meta combined supervised fine-tuning with on-policy distillation and reinforcement learning across general knowledge, reasoning, coding and agentic tasks.
The company said Muse Glimmer was assessed under its Advanced AI Scaling Framework before being approved for an open-weight release.
Muse Glimmer Benchmark Results
Meta compared Muse Glimmer with Google’s Gemma 4 31B and Alibaba Group’s (NYSE: BABA) Qwen 3.6 27B across agentic, coding, multimodal and reasoning benchmarks.
According to Meta’s published results, Muse Glimmer recorded the highest score among the three models on several tests, including MCP Atlas, DeepSearch QA, SWE-Bench Pro, SciCode and AIME 2026.
| Benchmark | Muse Glimmer 30B | Gemma 4 31B | Qwen 3.6 27B |
|---|---|---|---|
| MCP Atlas | 75.5 | 54.2 | 62.5 |
| DeepSearch QA | 74.6 | 61.7 | 71.1 |
| SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| SciCode | 43.6 | 43.4 | 39.8 |
| AIME 2026 | 94.7 | 89.2 | 94.1 |
| AA-LCR | 80.0 | 68.3 | 73.3 |
| Beam 128K | 65.1 | 58.2 | 63.0 |
Muse Glimmer did not lead every benchmark. Qwen achieved higher scores on tests including GDPval-AA, OSWorld-Verified, SWE-Bench Verified, TerminalBench 2.1, ScreenSpot Pro and OmniDocBench. Gemma recorded the strongest result on GPQA Diamond and Humanity’s Last Exam in Meta’s comparison.
These figures were published by Meta and should be treated as company-reported benchmark results rather than independent evaluations. Performance in real-world applications may differ depending on prompts, quantization, agent scaffolding, hardware and software configuration.
Muse Glimmer Supports Coding and Tool Use
Muse Glimmer is designed to perform more than standard text generation.
Meta says the model can invoke functions using defined schemas, maintain plans across multi-stage tasks and respond to tool failures by diagnosing problems and attempting another approach.
Its intended uses include:
- Local coding assistants
- File and document organization
- Personal scheduling agents
- Message-drafting tools
- Research and search workflows
- Software testing and debugging
- Automated model evaluation
- Visual document and screenshot analysis
Because these applications may require access to files, messages, credentials and other private information, local execution could give developers more control over where data is processed.
Local operation does not automatically make an agent secure. Developers must still manage permissions, untrusted inputs, tool access and stored information carefully.
Llama.cpp, MLX and ExecuTorch Support Planned
Meta said optimized integrations for llama.cpp, MLX and ExecuTorch would become available following the model’s release.
MLX is an array framework developed by Apple (Nasdaq: AAPL) for machine-learning research on Apple silicon. ExecuTorch is Meta’s framework for running AI models on edge devices, while llama.cpp is widely used to operate quantized models on consumer hardware.
These integrations are intended to make Muse Glimmer easier to deploy across Macs, PCs and other supported devices.
Developers can access the model weights, documentation and usage policies through Meta’s Muse Glimmer developer page.
Meta Expands Its Open-Weight AI Strategy
Muse Glimmer gives Meta a locally deployable model focused specifically on AI agents, coding, multimodal analysis and extended tool-based workflows.
The Apache 2.0 license allows developers to use, modify and redistribute the model weights, including for many commercial applications, subject to the license terms and Meta’s applicable usage policy.
For Meta investors, the launch highlights the company’s effort to expand its position in open-weight AI while supporting applications that run outside centralized cloud platforms.
For developers, Muse Glimmer’s appeal will depend on whether its benchmark performance translates into reliable real-world agents that can operate quickly and securely on accessible hardware.
