Skip to content
NewsModel release

Meta open-sources Muse Glimmer, a 30B agentic model that runs on a single consumer GPU

· by Pondero Newsdesk

The short version

Meta released Muse Glimmer on August 10, 2026 under Apache 2.0, the first open-weight model distilled from a frontier agentic system that fits in a 24 GB consumer GPU, with tool use, multi-step reasoning, and failure recovery built in.

Meta open-sources Muse Glimmer, a 30B agentic model that runs on a single consumer GPU

The gap between local inference and cloud-hosted agent reliability was real and measurable: agentic capabilities existed only in cloud APIs or in models requiring server-class hardware to run. Meta closed it on August 10, 2026, releasing Muse Glimmer under Apache 2.0 as the first open-weight model in its class to ship with native tool calling, multi-step reasoning, and automatic failure recovery in a package that fits on a single 24 GB consumer GPU. The 30-billion-parameter model is distilled from Muse Spark, Meta's internal frontier agent.

What

Meta trained Muse Glimmer on outputs from Muse Spark using logit distillation, per the Meta AI Research blog. Three training phases shaped the final model: pre-training on Muse Spark's outputs, mid-training on longer-context and agent-heavy data with richer reasoning traces, and post-training that combined supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains.

The model is a dense, decoder-only multimodal transformer with a built-in perception encoder that handles interleaved text and image input. A 128K-token context window is the default, with longer contexts supported, per Meta's developer documentation.

At full precision, a 30B model requires over 55 GB of memory, far outside consumer hardware. Meta addressed this with two optimizations. First, 4-bit quantization compresses the weights to under 20 GB in the K-Quant-17GB variant. That leaves enough headroom for the model's key-value cache, the perception encoder, and a speculative decoding component to run simultaneously within a 24 GB or 32 GB GPU envelope. Second, a lightweight companion network called DFlash proposes blocks of tokens in parallel, which the main model then verifies. Meta measured DFlash's effect on the K-Quant-17GB model: on an NVIDIA RTX 5090, DFlash produced 3.1 times higher throughput compared to standard token-by-token generation; on an Apple M5 Max, 1.8 times; on an M4 Max, 1.5 times. All three figures are from the research blog.

Agentic capabilities built into the model include reliable tool calling across extended multi-step workflows, multi-step reasoning across long horizons, and failure recovery: when a tool call fails or returns an unexpected result, the model diagnoses the error and retries rather than halting. The model also works with OpenClaw and other agentic orchestration scaffolds.

Meta benchmarked Muse Glimmer against Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety, and reasoning benchmarks including DeepSearch QA, MCP-Atlas, tau-Bench, and SWE-Bench, reporting strong performance in its size class. Those results are Meta's own evaluation. No independent third-party benchmark was available at the time of publication.

Weights are on Hugging Face now. Optimized integrations with llama.cpp, MLX, and ExecuTorch are scheduled for the coming days. Inference deployment partners at launch include Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter.

Why it matters

Muse Glimmer is the first open-weight model claiming to close the agentic reliability gap at the 30B tier, on hardware that costs under $2,000 at retail. Failure recovery, concretely, means the model diagnoses a broken tool call and retries within the same workflow pass, rather than halting and returning an error to the orchestrator.

For operators building with Cursor, Cline, or Aider against local inference servers, Muse Glimmer is the first model in this parameter class that shipped with agentic reliability benchmarks. The DFlash speedup figures mean agent loops that previously felt too slow for real work may become practical on a single-GPU workstation or high-end laptop.

The Apache 2.0 license is the broadest license Meta has applied to a frontier-class model. Commercial use, fine-tuning, and redistribution are all permitted without restrictions tied to user count or revenue. Operators who previously avoided open models because of license complexity have fewer objections here.

The privacy angle is direct. Agents that manage personal files, drafts, and schedules need access to sensitive personal context. Keeping that inference local, rather than routing it through cloud infrastructure, means the context stays on the device. Meta's stated design goal for Muse Glimmer was "always-on local agent workflows" and personal intelligence use cases.

One open risk: Meta has not disclosed the composition of Muse Spark's training data, which served as the teacher in the distillation process. Apache 2.0 says nothing about the underlying training pipeline, so the model's potential exposure to proprietary code or copyrighted content in the training set is currently unverifiable. Enterprise legal teams evaluating Muse Glimmer for production deployment should treat that gap the same way they treat it for other large open models where training data is undisclosed.

Context and reactions

Meta has a long record of open model releases, from early Llama versions through Code Llama and the Muse Spark family. Apache 2.0 on Muse Glimmer is the most permissive license in that sequence, removing the headcount and revenue restrictions that appeared in previous Llama-family licenses.

The two comparison models in Meta's evaluation, Gemma4-31B from Google and Qwen3.6-27B from Alibaba, both target efficient deployment in the 30B tier but were not specifically designed and benchmarked around agentic task completion. Meta's positioning is that Muse Glimmer occupies the same memory class while prioritizing the agentic workloads those models were not optimized for. The Hugging Face Open LLM Leaderboard is the first place community benchmarks will surface; tau-Bench scores run outside Meta's own evaluation suite will give the clearest read on whether the agentic positioning against Gemma4-31B and Qwen3.6-27B holds.

TechCrunch tied the release to Zuckerberg's stated interest in "personal intelligence" products and described the model as a signal of Meta's direction on device-resident AI assistants. Meta has not disclosed how Muse Glimmer feeds into specific consumer product timelines.

What to watch next

Community quantization variants beyond Meta's K-Quant-17GB and K-Quant-Dynamic will appear on Hugging Face within days. The practical question for local-inference operators is whether DFlash throughput gains hold on RTX 4090 hardware (16 GB to 24 GB VRAM configurations more common than the RTX 5090) and whether the agentic reliability figures replicate in real workloads outside Meta's own evaluation suite.

The training data disclosure gap is the slower-moving variable. The specific trigger to watch: a Muse Spark model card from Meta detailing training data composition, or a published third-party audit of the distillation pipeline. Either would change the IP-indemnification calculus for enterprise teams currently treating Muse Glimmer as another undisclosed-data open model.

Sources