Meta has released Muse Glimmer, a 30-billion-parameter open-weight AI model designed to run locally on consumer hardware while handling coding, tool use, multimodal input, and long-running agentic tasks.
The model comes from Meta Superintelligence Labs and is available under the permissive Apache 2.0 license. Meta has published the full weights and official quantized builds on Hugging Face, giving developers the option to run, modify, and deploy it without relying on a cloud API.
Muse Glimmer is a dense causal transformer with about 29.6 billion parameters, including a dedicated perception encoder for images. It supports a context window of more than 131,000 tokens and was trained for multi-step reasoning, reliable function calling, failure recovery, coding, and other AI agent workflows.
The main attraction is hardware efficiency. Meta says quantization reduces the language model to under 20 GB at roughly 4-bit precision. Its smallest official GGUF build is about 17 GB and targets machines with 24 GB of VRAM, leaving additional memory for the KV cache and optional components used for vision and faster decoding.
Muse Glimmer Is Built to Stay on the PC
Meta positions Glimmer as an alternative to powerful models that normally require remote data centers. The model can run completely locally without sending prompts, files, code, or images to an external inference service, which could appeal to developers and companies working with private repositories or sensitive local data.
The official model page says Muse Glimmer is small enough to operate on a single consumer GPU or Mac. Meta tested its compact K-Quant-17GB version on an NVIDIA RTX 5090 and Apple M4 Max and M5 Max systems.
With Meta’s DFlash speculative decoding component enabled, the RTX 5090 reached an average of 233.4 generated tokens per second in the company’s test, up from 74.9 tokens per second without speculation. The M4 Max reached 37.8 tokens per second and the M5 Max 50.2 tokens per second with DFlash.
Those numbers do not mean every ordinary home PC can run the model comfortably. A 24 GB graphics card remains high-end consumer hardware, while machines with less memory may need heavier compression or partial CPU offloading. Still, putting a 30B multimodal agent model into a single-device memory envelope is a significant shift from the multi-GPU servers commonly associated with capable agentic systems.
AMD is also supporting the release. The chipmaker said Glimmer can run locally on Ryzen AI Max+ systems and a single Radeon AI PRO R9700 using llama.cpp, expanding the range of hardware beyond NVIDIA and Apple platforms.
Meta Distilled Glimmer From Muse Spark
Muse Glimmer was distilled from the larger Muse Spark model. Knowledge distillation uses a more capable teacher model to train a smaller system that retains useful behavior with substantially lower memory and inference requirements.
That relationship explains why Glimmer is focused less on being a general frontier model and more on a specific operating environment: autonomous agents running continuously on a personal machine. Meta says the model was trained to maintain plans across long workflows, make structured tool calls, recover when tools fail, and interpret both text and images.
Its benchmark results are strongest in the same areas. Meta reports a score of 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 51.2 on SWE-Bench Pro, and 76.0 on SWE-Bench Verified. On several of these tests, Glimmer matches or exceeds other models in the roughly 30-billion-parameter class, though the results are Meta’s own evaluations rather than independent testing.
Meta also says the compact K-Quant-17GB build loses only about 1% on average across 15 benchmarks compared with full precision. Its larger dynamic quantization targets 32 GB of VRAM and is reported to lose about 0.2%.
The release follows Meta’s recent expansion into coding agents. The company launched Muse Code in early August as a terminal-based agent powered by Muse Spark 1.2 for long-running software projects. Glimmer takes the same agent-first direction but moves the model onto the user’s own machine.
Open Weights Return to the Center of Meta’s AI Strategy
Muse Glimmer also marks a renewed open-weight push from Meta after the company became more cautious about releasing its most advanced systems. CEO Mark Zuckerberg said more open models are coming and argued that advanced AI should not remain concentrated inside a small number of companies and governments.
Reuters reported that Meta also plans to release the weights of Muse Spark 1.2, its more capable model, as the company tries to regain momentum in open AI against rapidly improving Chinese model developers.
For local AI, however, Glimmer may be the more immediately practical release. A frontier-scale model that requires a data center remains inaccessible to most developers regardless of whether its weights are downloadable. A 30B model that fits on one powerful workstation can be experimented with, fine-tuned, integrated into local tools, and run continuously without token fees.
That is the bet behind Muse Glimmer. Meta is not trying to make the smallest possible model. It is trying to put enough reasoning, coding, vision, and tool-use capability into a package that a serious developer can keep running on a desk rather than in someone else’s cloud.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
