Meta Releases Muse Glimmer: A 30B Open-Weight Agentic AI Model for Local Hardware

Meta released Muse Glimmer, a 30-billion-parameter open-weight model designed to run autonomous AI agents locally on consumer hardware. According to Meta, the model is optimized for high-end Macs and PCs with 24GB to 32GB of VRAM, allowing agentic workloads to operate without relying on cloud infrastructure.

The model is released under the Apache 2.0 open source license, which Meta’s technical blog and official announcements state permits unrestricted commercial use, modification, and redistribution. This represents a shift from the bespoke community license used for the Llama family, which included restrictions such as a 700-million-monthly-user cutoff.

Meta CEO Mark Zuckerberg announced the release on X, stating that the company would soon also release the weights for Muse Spark 1.2, the foundation model behind the Muse Code terminal coding agent.

Agentic Architecture and Hardware Requirements

Meta trained Muse Glimmer specifically for the operational loop of autonomous agents. According to Alexandr Wang, Meta’s chief AI officer, the model is designed to formulate plans, call tools, interpret results, and recover from failures. Wang stated on X that the model can run on 24GB of VRAM without losing agentic reliability.

The model is a dense causal transformer with approximately 29.6 billion parameters across 52 layers, including a 1.8B-parameter ViT-G/14 perception encoder, according to its Hugging Face model card. It supports interleaved text and images, over 100 languages, and a context length of at least 131,072 tokens, with a knowledge cutoff of January 4, 2026.

To fit the model on consumer GPUs, Meta developed 4-bit quantized versions. The company reports two primary configurations:

• K-Quant-17GB: Targeted at 24GB VRAM hardware, such as Nvidia RTX 3090 or 4090 GPUs.
• K-Quant-Dynamic: Targeted at 32GB VRAM hardware, such as the Nvidia RTX 5090 or Apple Silicon Macs with 32GB or more of unified memory.

Meta’s internal measurements claim an accuracy degradation of 1% for the 24GB configuration and 0.2% for the 32GB version across 15 benchmarks.

Inference Speed and Speculative Decoding

To reduce latency in local agent workflows, Meta implemented DFlash speculative decoding. This system uses a smaller drafter model to propose blocks of 16 tokens that the primary model verifies in parallel.

Meta reports that this increases generation speeds on an Nvidia RTX 5090 from 74.9 to 233.4 tokens per second. On Apple hardware, speeds rose from 23.7 to 37.8 tokens per second on the M4 Max and from 26.6 to 50.2 tokens per second on the M5 Max, according to company data.

Competitive Benchmarks and Local Model Market

Muse Glimmer competes with other open-weight models in the 30B size class, including Google’s Gemma 4 and Alibaba’s Qwen3.6-27B. In Meta’s own evaluations, Glimmer outperformed these models on several agentic tests, including MCP Atlas (75.5), DeepSearch QA (74.6), and SWE-Bench Pro (51.2).

However, Meta’s data shows Qwen3.6-27B leads in other areas, specifically on OSWorld-Verified (75.6 vs. Glimmer’s 65.9) and TerminalBench 2.1 (60.7 vs. 51.7). Gemma 4 leads on GPQA Diamond and Humanity’s Last Exam.

The release follows a trend of high-capability open-weight models from Chinese labs. According to OpenRouter usage data from May 2026, Chinese open-weight models accounted for roughly 61% of all tokens consumed on the platform.

Safety and Deployment Guardrails

The model weights are available on Hugging Face. Meta stated that support is rolling out through Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter, with further integrations for llama.cpp, MLX, and ExecuTorch arriving in the coming days.