Meta Muse Glimmer Brings Apache 2.0 AI Agents to 24GB PCs
Meta's Apache 2.0 Muse Glimmer brings multimodal AI agents to 24GB and 32GB computers, with open weights, fast decoding and local privacy for developers.
Summary
Meta released Muse Glimmer on August 10, 2026, a 29.6 billion parameter dense multimodal model built for autonomous agents on high end Macs and PCs. Its Apache 2.0 weights permit unrestricted commercial use, modification and redistribution, Meta's first fully open release since proprietary Muse Spark replaced Llama in April and more permissive than Llama's license with its 700 million monthly user cutoff. Hugging Face hosts BF16 weights, two 4 bit quantizations, the DFlash drafter and 1.8 billion parameter ViT-G/14 encoder. Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter support arrives this week, followed by llama.cpp, MLX and ExecuTorch, while Unsloth and TorchTitan support deployment and tuning.
Glimmer handles interleaved images and text, more than 100 languages and at least 131,072 tokens, with knowledge through January 4, 2026, four reasoning levels, tool use, self checking and failure recovery. Distilled from Muse Spark and trained across reasoning, coding and agent tasks, it supports OpenClaw and Hermes Agent. Full precision needs more than 55GB and BF16 is pegged at 64GB, but 4 bit builds fit the complete agent stack into 24GB or 32GB. Meta reports 1% and 0.2% average accuracy loss across 15 benchmarks. DFlash raised RTX 5090 throughput from 74.9 to 233.4 tokens per second, M5 Max from 26.6 to 50.2, and M4 Max from 23.7 to 37.8.
Meta's tests put Glimmer ahead of Gemma 4 31B and Qwen3.6-27B on several agent benchmarks and at 51.2 on SWE-Bench Pro, but Qwen led several computer, terminal and multimodal tests. Meta rated Glimmer Moderate or lower risk and recommends guardrails plus human confirmation for irreversible actions. Zuckerberg said Muse Spark 1.2, the foundation model behind Muse Code released five days earlier, will follow. Local inference can protect sensitive data and remove network and token fees, but hardware, electricity and management costs remain.
Positives
- K-Quant-17GB runs on 24GB RTX 3090 or RTX 4090 cards, while K-Quant-Dynamic targets the 32GB RTX 5090.
- Meta's Home Assistant demo discovered devices, queried APIs, built an HTML, CSS and JavaScript dashboard, deployed it locally and checked the result.
- Glimmer scored 75.5 on MCP Atlas, 74.6 on DeepSearch QA, 23.5 on τ³-Banking, 47.6 on WildClawBench and 43.3 on GAIA2.
- AMD, Arm, Dell, Intel and Nvidia are working with Meta on device optimization.
Risks & concerns
- Typical 8GB and 16GB laptops cannot run Glimmer's quantized builds, limiting local deployment to expensive high end systems.
- Qwen3.6-27B beat Glimmer 75.6 to 65.9 on OSWorld-Verified, 60.7 to 51.7 on TerminalBench 2.1 and 77.2 to 76.0 on SWE-Bench Verified.
- Glimmer recorded 26.4 on CI Memories and 28.4% prompt injection success on Siren AgentDojo, despite its 94.2 utility score.
- Meta's quantization and speed claims are company measurements, with real repository, enterprise tool and long session performance still unproven.
- By May 2026, Chinese models generated roughly 61% of OpenRouter tokens, four of its five leading models were Chinese, and Llama had left the rankings.
- Meta opened Glimmer's weights, quantizations, drafter and perception encoder, but not its training data or training code.