Official GGUF
Available
Meta publishes the K-Quant files and companion artifacts in the official repository.
Inspect official filesUPDATED AUG 10, 2026 · SOURCE-BACKED
Muse Glimmer 30B: Can You Run It Locally?
Start with the published artifact sizes, Meta's capacity targets, and a runtime-aware KV range. The estimator answers hardware fit without pretending a target tier is a tested machine.
- Architecture
- 29.6B dense
- Config limit
- 131K context
- Inputs
- Text + Image
- Terms
- Apache-2.0 + Usage Policy
Facts checked against the Meta introduction, official model card, and Usage Policy.
Hardware-fit estimator
Can I run it?
Result
Official target
Estimate, not a guarantee
Meets Meta's published capacity target for this variant. This is a target tier, not verified device compatibility.
Active allocation
21.63 GiB active / 26.43 GiB conservative
Capacity 24 GiB
Meta target 24 GiB
- Weights15.61 GiB
- Vision1.3 GiB
- DFlash1.52 GiB
- Active KV1.7 GiB
- Planning reserve1.5 GiB
- Headroom2.37 GiB
- Active KV
- 1.7 GiB
- Conservative KV
- 6.5 GiB
- Active headroom
- 2.37 GiB
- Conservative headroom
- -2.43 GiB
How this range is planned
The site adds a 1.5 GiB planning reserve for runtime overhead. It is a site assumption, not an official requirement.
Active KV follows the model's global/local attention pattern. The conservative total assumes every layer allocates the full selected context, so actual runtimes can land between the two estimates.
Evidence: Meta published target tier
Official target means Meta's published capacity tier, not a verified device result.
- Capacity arithmetic uses binary GiB (1 GiB = 1,073,741,824 bytes).
- 1.5 GiB is reserved for runtime overhead as a site planning assumption, not an official requirement.
- Active KV uses 13 global layers at the selected context and 39 local layers capped at 2,048 tokens.
- KV arithmetic uses 16-bit KV (2 bytes per element) with 2 KV heads and head dimension 128; common F16/BF16 cache allocation is covered, but runtime cache type and allocation may vary.
- Conservative KV assumes all 52 layers allocate the full selected context.
Published capacity tiers
Official targets
| Memory target | Variant | Meta quality note | Planning note |
|---|---|---|---|
| 24 GiB | K-Quant 17 GB | 1% average degradation | Meta target for the smaller official quant; task-level quality can differ from the reported average. |
| 32 GiB | K-Quant Dynamic | 0.2% average degradation | Meta target for the dynamic official quant; additional memory can absorb runtime-dependent KV allocation. |
| 64 GiB | BF16 | Reference weights | Meta target for the official BF16 weights; this capacity tier is not a verified device result. |
Runtime watch
What is verified right now
llama.cpp
Merged to master
Muse Glimmer support was merged at commit 62bf73d. Older builds may fail before model loading or lack the required support.
Other integrations
Unknown
Version-specific compatibility for integrations not backed by a dated source has not been verified here. Availability is not inferred.
Method and limits
What the estimate can - and cannot - tell you
The calculator adds exact published file sizes, selected companion artifacts, a KV-cache estimate derived from the model configuration, and a small site planning reserve. It reports active and conservative totals because allocation behavior depends on the runtime.
It is a capacity screen for choosing what to investigate next. It does not measure speed, prove that a particular GPU or Mac works, or replace build-specific runtime documentation.
- 01
Estimate, not a guarantee. No device is certified by this tool.
- 02
Target, not minimum. Official target means Meta's capacity tier.
- 03
Reserve is local policy. The 1.5 GiB runtime reserve is this site's planning assumption.
- 04
KV varies. Runtime cache type and allocation can move the real total.
- 05
Support is versioned. Unsourced integration availability remains unknown.
Continue the local setup
- Muse Glimmer GGUF guide
Compare the two official K-Quant artifacts and companions.
- System requirements
Plan memory, context, storage, and runtime reserve.
- llama.cpp setup
Follow source-dated runtime support and setup guidance.
- Troubleshooting
Diagnose loading, allocation, and build-version errors.
- Benchmarks
Separate published quality claims from measured local results.