Guide · Local AI
Muse Glimmer hardware requirements
Meta released Muse Glimmer on August 10, 2026 as a 30B model for local agent workflows. The official announcement links the weights and developer resources.
Use the official artifact names
The Meta model card identifies these targets:
| Release | Publisher hardware target |
|---|---|
| K-Quant-17GB | 24GB VRAM |
| K-Quant-Dynamic | 32GB VRAM |
| Full precision | 64GB VRAM |
These are publisher targets, not our measured minimums. Community Q3, Q6 and Q8 artifacts need their own sizing and quality evidence. NVFP4 and GGUF K-quants are not interchangeable runtime formats.
What the published speedup actually measures
Meta reports 3.1× decoding speedup on an RTX 5090 with its K-Quant-17GB artifact and DFlash drafter, versus no speculation on that same hardware. The reported test used batch size one and greedy decoding through llama.cpp. This is not a 5090-versus-3090 benchmark. Mac results used ExecuTorch, so they should not be advertised as MLX results.
Check the complete runtime
The NVIDIA-hosted model card describes a specific vLLM integration with listed Blackwell and Hopper hardware. Its support list is not evidence that the same deployment works on Ampere or AMD. Confirm the release recipe for your own backend.
- Choose the exact artifact and supported runtime version before shopping.
- Account for the perception encoder and drafter if you enable them, plus context and working memory. The advertised context limit is not a guarantee it fits your device.
- On Macs, allow for macOS and other apps. Do not equate installed unified memory with available VRAM.
- Test representative prompts and tool calls at your intended context. A smaller quantization can change results; no bit depth guarantees correct tool use.
Hardware to investigate
The 3090 and 4090 provide 24GB; the 5090 provides 32GB. These are capacity candidates, not configurations we have tested. Use their dated price samples after confirming a compatible recipe. Larger workstation cards and high-memory Macs deserve the same check.
We are withholding a “cheapest guaranteed setup” verdict until there is reproducible evidence for the exact model artifact, runtime, context and hardware. Old fair-value estimates cannot establish the cheapest machine available today.
FAQ
How much memory does Muse Glimmer need?
Meta lists 24GB for K-Quant-17GB, 32GB for K-Quant-Dynamic and 64GB VRAM for full precision. These publisher targets do not establish universal minimums for other artifacts or runtime settings.
Should I buy a 5090 instead of a 3090 for Muse Glimmer?
The 5090 offers 32GB rather than 24GB. Confirm that its extra capacity or a matching workload benchmark justifies the current price. The published speculative-decoding speedup is not a comparison against a 3090.
Sources and review
Hardware advice reviewed October 4, 2026. These are documentation-based recommendations; we have not benchmarked these configurations.
- Ollama hardware and operating-system support
- Ollama context and memory guidance
- MLX LM model support and memory controls
- llama.cpp multi-GPU configuration
Compare GPU prices · Estimate memory needs · Compare Mac and GPU builds