Meta's Muse Glimmer: The 30B Open-Weight AI Model That Runs Entirely on a Single Consumer GPU

Key Specifications

Muse Glimmer is a dense causal transformer (not Mixture-of-Experts) with a dedicated ~1.8B-parameter ViT-G/14 perception encoder for multimodal input . It features 52 layers, a hidden dimension of 6,656, and 32 query heads with a 16:1 Grouped Query Attention pattern .

The full-precision model requires ~55 GB of VRAM, but Meta ships it with 4-bit quantization that compresses the model to under 20 GB .

Quantized VRAM Variants

Meta ships two quantized release configurations, both built with K-Quant and measured against 15 benchmarks :

• K-Quant-Dynamic — targets ~32 GB VRAM, with only 0.2% average benchmark degradation vs. full precision
• K-Quant-17GB — targets ~24 GB VRAM (model is under 20 GB), with 1.0% average degradation

Independent laboratories have not yet reproduced these aggregate degradation claims .

DFlash Speculative Decoding & Local Throughput

To solve the latency problem of running a 30B model locally, Meta pairs Muse Glimmer with DFlash — a small five-layer speculative-decoding drafter that proposes blocks of 16 tokens at once, which the main model then verifies in parallel .

Meta reports the following throughput on the 17 GB quantized model at batch size 1 with greedy decoding :

SGLang reported 236.4 tok/s on the same RTX 5090 hardware in a separate day-0 measurement . Early independent tests on an RTX 5090 with llama.cpp sometimes report lower numbers (e.g., ~137 tok/s), suggesting that the 233+ tok/s figure requires specific optimizations .

Agentic Capabilities

Muse Glimmer is not primarily a chatbot — it is a local AI agent designed for autonomous workflows . Meta explicitly tuned it to plan, call tools, and recover from its own errors . It supports the Model Context Protocol (MCP) natively and integrates with popular agent frameworks .

Key use cases include: function calling, local coding, long tool-use sessions, LLM-as-a-judge evaluation, document analysis, and personal assistants .

Benchmark Performance vs. Peers

Meta published direct comparisons against Google's Gemma 4 31B and Alibaba's Qwen 3.6 27B (both in thinking/reasoning mode) :

Muse Glimmer leads on most agentic benchmarks, including MCP Atlas (75.5 vs. 62.5 for Qwen) and DeepSearch QA (74.6 vs. 71.1) . It also scores 94.7 on AIME 2026, leading both competitors .

Caveat: One third-party analysis notes that Glimmer trails Qwen 3.6 27B on TerminalBench 2.1 (51.7 vs. 60.7) and OSWorld-Verified (65.9 vs. 75.6), and has a tendency to refuse certain OS-level automation tasks .

Distillation from Muse Spark

Muse Glimmer was distilled from Muse Spark (Meta's larger, closed-source frontier model) using logit distillation during pre-training, followed by supervised fine-tuning and RLHF . Muse Spark itself remains closed — available only as a hosted API — making Glimmer the open-weight distillation for users who want local deployment . Meta has indicated that an open release of Muse Spark may follow .

Strategic Context: Superintelligence for Everyone

Muse Glimmer is the clearest product yet of Mark Zuckerberg's vision of "personal superintelligence" — powerful AI running locally on personal devices, free from cloud costs, subscription fees, or data leaving the device . TechCrunch called it "the clearest picture yet" of that vision .

The release follows months of internal turmoil after Llama 4's disappointing launch in spring 2026, after which Meta appeared to abandon open-weight AI. Muse Glimmer reopens that strategy .

The $14.3B Scale AI Investment

In June 2025, Meta invested $14.3 billion in data-labeling firm Scale AI, acquiring a ~49% stake . The deal brought Scale's 28-year-old CEO Alexandr Wang to lead Meta's new "Superintelligence" division . The investment was aimed at securing high-quality training data and talent for the Muse model family . CNBC reported in January 2026 that Wang was leading Meta's TBD AI unit building a successor to Llama, code-named Avocado .

Availability and Day-0 Support

Muse Glimmer is downloadable from Hugging Face at meta-models/Muse-Glimmer-30B under the Apache 2.0 license . It has day-0 support in:

• Ollama — pull and run locally with a single command
• LM Studio — supported for local GUI-based inference
• SGLang — 230+ tok/s on RTX 5090
• llama.cpp — community GGUF conversions available day-of
• NVIDIA NIM — optimized inference via NVIDIA's microservice framework
• AMD ROCm — official support for Ryzen AI Max and Radeon GPUs

Early community reports show that even an RTX 3090 (24 GB) can run the Q4_K_XL quantized model with DFlash and a 262K context window using about 22-23 GB of VRAM .