Launched

Meta’s New Open AI Model “Muse Glimmer” Runs on Your Computer, of Course With Trade-offs

co-written by eustella AI10. August 2026, 13:17

Meta Superintelligence Labs introduced “Muse Glimmer” on Monday: a 30-billion-parameter model that, according to the company, can run on a Mac or a PC with a single consumer GPU. The weights are available on Hugging Face under the Apache 2.0 license – considerably more permissive than with earlier Llama models. In return, expectations around performance need to be recalibrated against the large cloud models.

The model’s purpose is narrowly defined. According to Meta, Muse Glimmer is built for “always-on” agent workflows that run permanently on local hardware. The company lists local agents and function calling, local coding, and use as an “LLM-as-a-judge” – an evaluating instance for other models’ outputs – as its application areas. The reasoning behind it: an agent that manages calendars, drafts messages and organizes files needs deep access to personal context, and that context does not necessarily have to travel to the cloud. Offline operation without an internet connection comes on top of that.

How the model fits onto a laptop

Meta did not train Muse Glimmer from scratch but distilled it from a larger model. In pre-training, according to the company, Muse Glimmer learned from the outputs of the larger Muse Spark model via logit distillation. Mid-training followed with longer contexts, more agent-heavy data and richer reasoning traces; post-training combined supervised fine-tuning, on-policy distillation and reinforcement learning.

Two technical interventions were required to make the result fit on consumer hardware, both of which Meta describes openly:

Quantization. At full precision, a 30-billion-parameter model would require more than 55 GB of memory according to Meta – more than any consumer GPU offers. Compressing the weights to roughly 4-bit precision shrinks the language model to under 20 GB. That leaves room within a 24 GB or 32 GB budget for the KV cache, the perception encoder for image understanding, and the draft model. Meta writes that the compression causes “minimal to no” degradation on agentic tasks.
Speculative decoding. A lightweight “drafter” model based on DFlash proposes entire blocks of tokens at once, which the main model verifies in parallel. Meta puts the speedup at 3.1x on an RTX 5090, 1.8x on a MacBook with M5 Max, and 1.5x on an M4 Max.

Functionally, the model covers the usual agent requirements: tool calls with precise schemas, multi-step reasoning over longer horizons, error handling when tool calls fail, multimodal input combining text and images (screenshots, charts, documents), adjustable reasoning strength, and training on data from more than 100 languages. As a scaffold, Meta names OpenClaw among others.

Where the limits are

Anyone comparing Muse Glimmer to the leading cloud models needs to keep several constraints in mind – some of which follow from the announcement itself.

The comparison is missing. Meta benchmarks Muse Glimmer exclusively against models of a similar size, specifically Gemma4-31B and Qwen3.6-27B. The blog post’s wording is that the model performs strongly “for its size class.” No benchmark comparison against the large proprietary cloud models is drawn. Meta names DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench as test suites. All figures so far come from the vendor itself; independent measurements, for instance by Artificial Analysis, are still outstanding.

Size still matters. Thirty billion parameters stand against frontier models that deploy a multiple of that in mixture-of-experts architectures. For narrowly defined, recurring agent tasks with clear tool schemas that may well be enough – but on broad world knowledge, difficult reasoning and niche domains, the gap upward is structural.

Quantization costs precision. Meta’s claim that compression to 4 bits barely degrades results refers explicitly to agentic tasks. No corresponding figure is given for other task types.

Context competes with the model for memory. Meta does not state a context length. On local hardware, the KV cache shares the memory budget with the model itself – which makes long contexts more expensive than in the cloud, where they simply scale along.

Speed remains relative. Meta gives only speedup factors, no absolute tokens per second. And the reference devices named – MacBooks with M4 Max or M5 Max, an RTX 5090 – sit at the upper end of the consumer segment. “Runs on a Mac or PC” means, in practice: 24 to 32 GB of graphics or unified memory.

Operations and security sit with the user. Running locally means no server-side updates, no downstream moderation layers and no availability guarantees. Meta states it evaluated Muse Glimmer under its in-house “Advanced AI Scaling Framework” and cleared it for open-weight release. With open weights, however, it generally holds that safeguards can be removed again through fine-tuning.

What the Apache 2.0 license permits

The licensing side is the actual novelty. Apache 2.0 is a permissive license recognized by the Open Source Initiative. In concrete terms:

Permitted are commercial use, modification, redistribution and sublicensing – with no revenue or user caps.
No copyleft: anyone modifying the model or building it into a product does not have to disclose the changes. Derivative works may remain proprietary.
Explicit patent grant: contributors grant rights to their patents. Anyone who files a patent lawsuit over the work, however, forfeits that license (retaliation clause).
Obligations: the license text and copyright notices must be included, changes to files must be marked, and any existing NOTICE file must be passed along.
Not included are trademark rights – Meta’s names and logos may not simply be used for one’s own products. Warranty and liability are disclaimed.

This removes the restrictions that characterized earlier Llama models. Their community license was not an open-source license in the narrow sense but contained, among other things, a threshold of 700 million monthly active users above which a separate agreement with Meta was required, an attached acceptable use policy, and naming requirements.

One caveat remains: Apache 2.0 applies to the released weights. Training data and the full training stack are not disclosed. Muse Glimmer is therefore an open-weight model under an open-source license – it is not reproducible in the sense of a complete rebuild.

Availability

The weights are available on Hugging Face. Integrations for llama.cpp, MLX and ExecuTorch are set to follow in the coming days; the model can be run locally via Ollama, LM Studio and Unsloth, while for larger-scale deployment Meta names vLLM and SGLang as well as the providers Together AI, Fireworks AI and OpenRouter. Customization runs through PyTorch’s TorchTitan. On hardware optimization, Meta says it is working with AMD, Arm, Dell, Intel and Nvidia.