Muse Glimmer

Quick Take

Muse Glimmer is Meta Superintelligence Labs’ first fully open-weight model since it moved its flagship line to the proprietary Muse Spark, released under the Apache 2.0 license on August 10, 2026. It’s a 30-billion-parameter dense model built specifically for local, agentic workloads: tool use, coding, and multi-step reasoning, running entirely on a single consumer GPU rather than requiring cloud infrastructure.

Muse Glimmer, Muse Code, and Muse Image Are Three Different Products

Meta’s “Muse” name now covers three separate things, and it’s worth keeping them straight:

• Muse Code is a paid, closed terminal coding agent, powered by the separate Muse Spark 1.2 model.
• Muse Image is Meta’s image-generation model, built into the Meta AI app, Instagram, and WhatsApp.
• Muse Glimmer is an open-weight, downloadable agentic model you run yourself, not a packaged product with a subscription.

Glimmer is not a smaller version of Muse Code or Muse Spark. It’s a separate model with its own training process, aimed at a different job: running locally rather than being served from Meta’s infrastructure.

What Muse Glimmer Is Built For

Muse Glimmer supports tool use, vision input, and reasoning, which Meta positions for autonomous local agents, function calling, local coding assistance, and using the model itself as an evaluator (“LLM-as-a-judge”) for other AI output. All of that runs without depending on a cloud API once the weights are downloaded.

Key Facts

• Developer: Meta Superintelligence Labs
• Released: August 10, 2026, as an open-weight release
• License: Apache 2.0
• Architecture: Dense, 30 billion parameters (every parameter activates per token; no Mixture-of-Experts routing)
• Context window: 131,072 tokens (128K)
• Modalities: Text, vision (image input), tool use
• Weights: Hosted on Hugging Face at meta-models/Muse-Glimmer-30B, with GGUF and ExecuTorch PTE quantized variants also published
• Local hardware: Fits within the VRAM of a single consumer GPU (Meta and NVIDIA cite the GeForce RTX 5090), with additional support for DGX Spark, DGX Station, and Jetson platforms
• Commercial use: Permitted under the Apache 2.0 license

How It Was Trained

Meta describes a multi-stage process: Glimmer was pre-trained on Muse Spark’s own outputs using logit distillation, mid-trained on longer-context, agent-heavy data with richer reasoning traces, then post-trained with supervised fine-tuning, on-policy distillation, and reinforcement learning across general, reasoning, coding, and agentic domains. In practice, that means it’s distilled from a larger, more capable sibling model rather than trained from scratch at this size.

Dense, Not Mixture-of-Experts

It’s worth being explicit about this since it affects how Glimmer compares to other open models in this size range: Glimmer is a dense model, not a Mixture-of-Experts one. Every one of its 30 billion parameters activates on every token. That’s a different tradeoff than a sparse MoE model like GLM 5.3 or DeepSeek’s V4 family, which activate only a fraction of a much larger total parameter count per token. Dense models are simpler to reason about for local deployment, since VRAM needs are predictable and don’t depend on expert routing, but they don’t get the same per-token compute discount a well-designed MoE model can offer at a larger total size.

Quantized Versions and Local Tooling

Meta published GGUF quantized weights alongside the full-precision release, and Unsloth added day-zero support for running and fine-tuning Muse Glimmer through its local tooling. That combination, official quantization plus fast third-party support, is a meaningful part of why it’s positioned as practical for consumer hardware rather than only for teams with server-grade GPUs.

Why This Matters for Ainanza Readers

Muse Glimmer connects several ideas that matter for anyone thinking about running AI outside a vendor’s cloud: open models versus closed ones, local AI versus API dependence, what it actually takes to run a coding agent with real tool use on your own hardware, and what it means to own your model rather than rent access to one. It sits alongside other open agentic models like Qwen3.8-27B and GLM 5.3 as evidence that a genuinely capable local model no longer requires a datacenter.

Honest Limitations

• Dense architecture caps efficiency. At 30B parameters with no MoE routing, it doesn’t have the extreme per-token cost advantage a sparse model of similar total size can offer.
• Newest entrant, least track record. Meta’s earlier open releases had a mixed reputation; Glimmer’s real-world reliability on your own tasks is worth testing directly rather than assuming from the launch announcement.
• Still needs a real GPU. “Runs on a single consumer GPU” means something like an RTX 5090, not an integrated laptop chip. Quantized GGUF builds lower the bar but with a quality tradeoff.
• No official hosted API from Meta. Unlike Muse Code, Meta isn’t offering to run Glimmer for you; self-hosting or a third-party host is the only path.

Alternatives Worth Knowing

• Qwen3.8-27B, Alibaba’s open-weight dense model in a similar size and local-deployment niche
• GLM 5.3, Z.ai’s larger Mixture-of-Experts open-weight model for coding and agents
• DeepSeek, an open-leaning model family with a strong cost-efficiency track record
• Nativ, LM Studio, and Ollama, tools for actually running a model like Glimmer locally
• Meta Muse Code, Meta’s separate, closed coding agent product

For the bigger picture on this category, see Frontier Open Models Explained and Open Models vs Closed Models.