Meta Superintelligence Labs has introduced Muse Glimmer, a 30-billion-parameter artificial intelligence model designed specifically for running agentic workloads locally on consumer devices.
The model is being released with its weights under the permissive Apache 2.0 license, giving developers and researchers the freedom to download, modify, deploy and build applications around it. Meta says Muse Glimmer is optimized for always-on local agent workflows, with potential applications including personal AI assistants, function calling, local coding, file management and automated evaluation of other language models.
Unlike many powerful AI systems that depend heavily on cloud servers and an internet connection, Muse Glimmer is designed to operate directly on a Mac or PC equipped with a single consumer GPU. The approach could make sophisticated AI agents more accessible in situations where users want greater privacy, lower network dependence or the ability to work offline.
As part of its continued commitment to open AI research, Meta is releasing the Muse Glimmer model weights on Hugging Face, alongside developer documentation for building and running AI agents. The model is designed to work with tools developers already use, with optimized integrations for llama.cpp, MLX and ExecuTorch expected to arrive in the coming days. These integrations are intended to make it possible for developers to move from downloading the model to running a functional agent within minutes.
Bringing agentic AI to local devices
Foundation models have made significant progress in reasoning, code generation and tool use, but many of today's AI applications continue to rely on remote infrastructure. Meta's latest model is aimed at narrowing that gap by bringing more capable agentic functionality onto personal computers.
A local AI agent can potentially perform tasks without continuously sending information to a cloud service. This is particularly relevant for assistants that need access to personal information, documents, schedules and files.
Meta describes Muse Glimmer as being built for tasks that require several capabilities to operate together. These include long-horizon execution, precise tool calling, multimodal understanding, long-context memory and instruction following.
The company argues that advances in open-source AI have demonstrated that smaller models can deliver increasingly competitive results when they are trained and optimized for specific tasks. Muse Glimmer follows that philosophy by prioritizing agentic workloads rather than attempting to compete solely on model size.
How Muse Glimmer was trained
Developing a model capable of operating effectively as an agent requires more than simply improving its ability to generate text. An agent must understand a user's objective, create a plan, interact with external tools, interpret their responses and recover when something goes wrong.
Meta therefore designed Muse Glimmer around a training process intended to transfer sophisticated agentic capabilities from a much larger teacher model while keeping the resulting system compact enough for local hardware.
The training process consisted of three major stages.
Pre-training involved training Muse Glimmer on outputs from Muse Spark using logit distillation. Meta says the model used a data mix similar to that of its teacher.
Mid-training shifted the emphasis toward longer contexts and more agent-heavy material. The model was exposed to richer reasoning traces alongside organic data, helping it deal with more complex and extended interactions.
Post-training combined supervised fine-tuning with on-policy distillation and reinforcement learning. The training covered general tasks, reasoning, coding and agentic applications.
Meta says Muse Glimmer was also evaluated according to the standards established by its Advanced AI Scaling Framework before being assessed for open-weight release across relevant categories.
Designed around the needs of AI agents
Muse Glimmer's capabilities extend beyond conventional question-and-answer interactions. Meta says the model has been trained and evaluated on several abilities considered important for autonomous agents.
End-to-end task completion: Muse Glimmer is designed to handle tasks from beginning to end rather than simply responding to individual prompts. Meta reports strong results on benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which test areas such as tool-based workflows, coding, debugging and multi-turn task completion.
Reliable tool use: The model can work with function calls and tool schemas throughout extended workflows, allowing it to interact with external software and services.
Multi-step reasoning: Muse Glimmer is intended to maintain coherent plans across lengthy tasks, rather than losing track of the objective after several interactions.
Failure recovery: When a tool call produces an error or an unexpected result, the model is trained to diagnose the problem and retry instead of simply stopping.
Multimodal understanding: A dedicated perception encoder allows the model to process interleaved text and images. This means an agent can potentially work with screenshots, charts and documents as part of a broader conversation.
Scaffold compatibility: Muse Glimmer is designed to operate with OpenClaw and other agentic orchestration patterns, allowing developers to incorporate it into existing agent frameworks.
Controllable effort: Developers can select different reasoning strengths depending on whether they want to prioritize response quality or speed.
Multilingual support: Meta says the model was trained on data covering more than 100 languages, broadening its potential use beyond English-language applications.
Performance against models in its class
Meta evaluated Muse Glimmer across a broad collection of benchmarks designed to measure the capabilities required for autonomous and agentic behavior.
According to the company, Muse Glimmer performs strongly for its size class on several widely used large language model benchmarks when compared with models such as Gemma4-31B and Qwen3.6-27B.
The emphasis, however, is not simply on benchmark scores. Meta is positioning the model around practical agent performance, particularly its ability to reason over multiple steps, use tools, interact with software and recover from failures.
Making a 30-billion-parameter model fit on consumer hardware
One of the biggest challenges with running a model of this size locally is memory.
At full precision, Meta says a 30-billion-parameter model would require more than 55 GB of memory, placing it well beyond the capabilities of many consumer GPUs.
To address this, Meta uses quantization to compress the model's weights to approximately 4-bit precision. The resulting language model can be reduced to less than 20 GB, leaving additional memory for the model's working context, known as the KV cache, as well as the perception encoder and speculative decoding system.
Meta says the compressed model can operate within a 24 GB or 32 GB memory envelope and that its testing showed "minimal to no degradation on agentic tasks" as a result of the compression.
This optimization is particularly important because a local agent needs more than the language model itself. It may simultaneously need memory for conversation history, images, tool interactions and other components involved in completing a task.
Speculative decoding aims to make local AI faster
Memory efficiency is only part of the challenge. For an AI agent to feel useful, it must also respond quickly enough to support continuous interaction.
Muse Glimmer therefore incorporates speculative decoding. Instead of generating every token sequentially through the main model, the system uses a lightweight "drafter" model based on DFlash to propose blocks of tokens.
The main model then verifies those proposals in parallel. Correct predictions can be accepted, while incorrect ones are corrected by the larger model.
Meta says this approach allows Muse Glimmer to generate responses significantly faster than conventional token-by-token generation while maintaining identical output quality.
The company is also providing quantized versions of the drafter to reduce the additional memory required by the system.
Testing Muse Glimmer on Macs and NVIDIA GPUs
Meta says it measured the performance of its K-Quant-17GB version together with the quantized DFlash drafter on MacBook M4 Max, M5 Max and NVIDIA RTX 5090 hardware.
The company describes the resulting performance as fast enough for "fluid conversation and real-time agent interaction," with the entire workload running locally on the user's device.
If the claimed performance translates consistently across different hardware configurations, the model could make local agent systems considerably more practical for developers and advanced users who do not want every interaction to depend on a cloud-based AI service.
Developers can download the model now
Muse Glimmer's model weights are available through Hugging Face, alongside documentation intended to help developers begin building their own agents.
Meta says integrations with llama.cpp, MLX and ExecuTorch are expected to arrive in the coming days. The model is also expected to become available through platforms and tools including Ollama, LM Studio and Unsloth.
For developers interested in larger-scale deployments, Meta says Muse Glimmer can be served through vLLM and SGLang. Cloud AI providers including Together AI, Fireworks AI and OpenRouter are also expected to provide access.
Developers will additionally be able to customize the model for particular applications. Meta points to PyTorch's TorchTitan training feature as an option for further tuning and adapting Muse Glimmer to specific use cases.
Hardware companies join the optimization effort
Meta says it is working with hardware partners including AMD, Arm, Dell, Intel and NVIDIA to improve Muse Glimmer's performance across different devices.
The company is also publishing documentation covering deployment and responsible development, including guidance for creating custom agent scaffolds.
The broader goal is to reduce the technical barrier between downloading an open model and having a functional personal AI agent running locally.
A push toward more open, personal AI
Muse Glimmer represents Meta's continued push into open-weight artificial intelligence, but with a particular focus on agentic systems that can operate directly on users' hardware.
The distinction is important. A conventional local chatbot can answer questions without a cloud connection, but an effective agent needs to do considerably more. It must be able to reason through a task, use tools, interpret different forms of information, remember context and respond appropriately when something goes wrong.
By combining a 30-billion-parameter model with quantization, multimodal processing and speculative decoding, Meta is attempting to make those capabilities practical on consumer hardware.
The company says the release extends its "long track record of open AI research" into agentic AI and gives developers access to local agentic capabilities.
With the weights now being opened and integrations with several local-AI and deployment frameworks planned, the next test will come from developers and the open-source community. Their applications will ultimately determine how useful Muse Glimmer becomes beyond the benchmarks and technical specifications.