Muse Glimmer / Dense
A local agent needs more than a small download
New to model sizes or GPU memory? Start with weights, parameters and quantization →
Muse Glimmer is worth evaluating for local agents because the release includes quantized artifacts and a DFlash drafter alongside the main model. The practical comparison is a complete on-device loop: perception, tool action and recovery.
Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Evaluate it for
Private on-device agents using text and image inputs on a capable workstation.
Choose another path when
Copying the 17 GB quantized artifact label into an all-in memory budget, or treating a speedup as independent of drafter and workload.
Release context
The 29.6B dense model uses alternating local and full attention with only two configured KV heads. Meta’s published local-device work makes quantization and speculative decoding part of the release story.
The publisher card gives August 2026, without a day. It is grouped after dated August releases; that placement does not imply it was older.
Deployment starting points
Start with an exact artifact and an engine that supports it. A configuration below is evidence of a documented path; its device count does not promise a particular throughput or concurrency.
Publisher evaluated setup
32 GB per GPU
Publisher’s local-device stack
The card evaluates this device with a quantized drafter. The 17 GB label excludes full runtime state; do not transfer its speedup to another workload.
Meta’s quantized main model and drafter
Checked 2026-09-13. Read the documented setup →
This is the index’s reported tensor payload, in decimal GB. It excludes file headers and runtime memory. Mixed precision, conversion and host/device placement determine how much GPU memory the loaded model needs.
Inspect the exact index metadata →This estimates inference memory for a hypothetical uniform precision. It is useful for rejecting an allocation that is too small. It does not establish a working deployment or estimate training memory.
| Representation | Approximate weight floor | What remains to budget |
|---|---|---|
| 16-bit uniform | 59.2 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 8-bit uniform | 29.6 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
| 4-bit uniform | 14.8 GB | Cache, modality state, workspace, quantization metadata and uneven sharding |
Enter the total across the whole model replica, not per GPU: KV or recurrent cache for all concurrent requests, image/audio state and execution workspace. There is no universal default. We additionally leave 15% of each GPU unused.
Enter a request-and-buffer budget to calculate the aggregate capacity requirement.
A count of one does not prove the engine supports this GPU. A count above one does not prove the model shards evenly or communicates efficiently. Publisher-reported, rounded model count; not an exact tensor census. Multimodal components and packaging can change the artifact size.
Architecture in practice
Acceptance rate and verification cost decide the gain.
Count drafter memory alongside the target model.
Mechanism schematic based on the pinned model card and configuration. It explains a design principle; it is not a full implementation graph.
| Exact checkpoint | meta-models/Muse-Glimmer-30B |
|---|---|
| Text model type | muse_glimmer_text |
| Layers | 52 |
| Hidden width | 6,656 |
| Attention / KV heads | 32 / 2 |
| Head dimension | 128 |
| Routed / selected experts | Not verified / Not verified |
| Configured positions | 131,072 |
| Documented extension | Not recorded for this checkpoint |
| Inputs → output | text, image → text |
| License metadata | apache-2.0 |
39 sliding attention layers; 13 full attention layers. These counts describe the configured pattern.
Open weights do not imply unrestricted use. Read the applicable license terms linked from the pinned card.
A useful comparison
These are dense multimodal alternatives; compare tool reliability and complete runtime memory, not just model size.
14 Aug 2026
A current dense baseline before you scale out
27B reported total · text + image + video
2 Apr 2026
The dense control for Gemma’s MoE experiment
30.7B reported total · text + image + video
These are editorial comparison candidates. We have not run a matched quality or serving benchmark, so this is not a ranking.
Prepare to explain it
Explain draft acceptance, verification cost and why a fast token stream can still produce a slow or unreliable tool-using agent.
Attention layouts · Quantization · Courses and worked examples
For a deployment evaluation, record exact weights, precision, engine, device count, interconnect, prompt/output lengths and concurrency. Report task success, errors, TTFT, TPOT and useful throughput together.
Our assessment is editorial judgment based on the linked architecture and deployment evidence. Model facts are publisher-reported or attributed to runtime maintainers; no independent GPU benchmark was run.
AI Infra Interviews, “Muse-Glimmer-30B: release, architecture and deployment”, checked 2026-09-13. Preserve this date and the exact checkpoint when citing.
