Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Muse-Glimmer family

Muse-Glimmer-30B

Hugging Face ↗ Follow changes

meta-models Apache 2.0 out 3 months ago

29.8B parameters, images and text in, text out, 128K tokens of context. 3 providers serve it from $0.30 per million input tokens.

Verified 2026-10-25 Rank #96 in LLMs · score 99.4 method v1 Sources 4 Downloads 30d 204.9k · #250 of 619 tracked

Changed this week: 12 new quantized builds on the Hub (191 in all).2026-10-24 · 11 new quantized builds on the Hub (190 in all).2026-10-24

Can I run it?18GB at INT4Estimatefits a 24 GB cardmethodCommunity: Muse Glimmer 30B: 24/32 GB, DFlash, benchmarks vs Qwen3.6-27B — AI on Mac →
What does it cost?$0.30/1M incheapest of 2DeepInfrabf162026-10-25
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?204.9kdownloads, 30 days#250 of 619 we track · top 40%+39 likes, downloads −17% this week2,084 likesHugging Face2026-10-24

Worth reading

Picked by rule from 24 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 1,209 HN pts

Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

Meta releases Muse Glimmer, a 30-billion-parameter open model licensed under Apache 2.0 for local agentic workflows.

  • Muse Glimmer trains on Muse Spark's outputs using logit distillation while leveraging a similar data mix as the teacher model.
  • The model accepts interleaved text and images through a dedicated perception encoder to interpret screenshots, charts, and documents.
research.meta.ai 2026-08-10 1,209 HN pts · 638 comments 1.2k words about the Muse Glimmer family family
Latest · 2 months ago

Muse Glimmer is a memory hierarchy disguised as a 30B Transformer

Meta's Muse Glimmer is a 30B-parameter multimodal agent designed to run autonomously on consumer hardware without cloud infrastructure.

  • The 13 full-attention layers still perform quadratic attention work during prefill, meaning a 131K prefill is not equivalent to a 4K prefill.
  • Global layers operate on position-aware representations because every key and value has already passed through three RoPE-equipped local layers.
abstractextraordinary.com 2026-08-18 16 HN pts · 9 comments 4.1k words about the Muse Glimmer family family

24 best of 24 tracked

The 2 picks above, plus these 22. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleMuse Glimmersimonwillison.net · 2026-08-11 · 4 HN pts
  2. DiscussionMeta releases Muse Glimmer, 30B parameter open-weight edge modeltwitter.com · 2026-08-10 · 18 HN pts
  3. ArticleMuse Glimmer: Meta's open model built for always-on local agents. 30B parametersdeveloper.meta.com · 2026-08-10 · 12 HN pts
  4. ArticleMuse Glimmer (High) Intelligence, Performance and Price Analysisartificialanalysis.ai · 2026-08-10 · 2 HN pts
  5. ArticleMuse Glimmer: Benchmarks and analysis | Artificial Analysisartificialanalysis.ai · 2026-08-10
  6. ArticleProving Muse Glimmer in Zero-Knowledge at 85 Tok/S on an H100attestable.com · 2026-08-11 · 4 HN pts
  7. ArticleExperience of Running Muse Glimmer on Macholaclaw.ai · 2026-08-11 · 4 HN pts
  8. ArticleFastest Inference Meta Muse Glimmer 30B on Applebasecompute.co · 2026-08-12 · 2 HN pts
  9. ArticleZuck rekindles open weights Llama drama with Muse Glimmertheregister.com · 2026-08-10
  10. ArticleMuse Glimmer 30B Architecture Notes | Sebastian Raschka, PhDsebastianraschka.com · undated
  11. ArticleFast, on Device Agentic AI with Muse Glimmer on ExecuTorchpytorch.org · 2026-08-14
  12. ArticleRun OpenClaw with Muse Glimmer Locally on Macholaclaw.ai · 2026-08-12
  13. ArticleMeta open-sources Muse Glimmer: 30B agent model runs locally under 20GB VRAMainexusdaily.vercel.app · 2026-08-10
  14. ArticleMuse Glimmer 30B: An Architecture and Hands-On Look at Meta's New Mid-Range Model - DEV Communitydev.to · 2026-09-18
  15. ArticleQwen 3.8 27B vs. Nemotron 3.5 vs. Muse Glimmer on an RTX 5090kgptalkie.com · 2026-08-14
  16. ArticleMeta Muse Glimmer 30B: Full Review (September 2026) - AIToolsReviewaitoolsreview.co.uk · 2026-08-17
  17. ArticleMuse Glimmer from Meta Superintelligence Labs is now availableollama.com · 2026-08-10
  18. ArticleMuse Glimmer 30B: A Comprehensive Technical Analysis - Local AI Zonelocal-ai-zone.github.io · 2026-08-15
  19. ArticleMuse Glimmer 30B: 24/32 GB, DFlash, benchmarks vs Qwen3.6-27B — AI on Macai-on-mac.com · 2026-08-12
  20. ArticleMuse Glimmer: Meta’s 30B Model Built for Efficient Inferencekaitchup.substack.com · 2026-08-11
  21. ArticleMuse Glimmer 30B Hardware and Benchmark Guide | Wavectwavect.io · 2026-08-10
  22. ArticleMuse Glimmer 30B Review — Benchmarks, Pricing & AI Panel Verdict | TopReviewed.aitopreviewed.ai · 2026-08-09

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

2 providers from $0.30

2 providers with a live price, from $0.30 to $0.35 per million input tokens; the median is $0.32. One dot per provider, cheapest in teal.

$0.010$0.10$1.00$10.00DeepInfra: $0.30 in / $1.20 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/meta-models/Muse-Glimmer-30B?expand[]=inferenceProviderMappingTogether AI: $0.35 in / $1.50 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/meta-models/Muse-Glimmer-30B?expand[]=inferenceProviderMappingcheapest $0.30median $0.32
$0.010$0.10$1.00$10.00DeepInfra: $0.30 in / $1.20 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/meta-models/Muse-Glimmer-30B?expand[]=inferenceProviderMappingTogether AI: $0.35 in / $1.50 out per 1M · observed 2026-10-25 · https://huggingface.co/api/models/meta-models/Muse-Glimmer-30B?expand[]=inferenceProviderMappingcheapest $0.30 · median $0.32 (white tick)
#Cheapest 3In /1MOut /1MContextQuantSeen
1 DeepInfra $0.30 $1.20 128K bf16 Oct 25
2 Together AI $0.35 $1.50 128K unknown Oct 25
3 Featherless — — — — Oct 25

All 3 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 3 API providers; heavily run locally (191 quantized builds); a big base for builders (72 fine-tunes, adapters and merges).

  • As a service3DeepInfra, Featherless, Together AI
  • On own hardware191quantized builds
  • Built on7256 fine-tunes · 13 adapters · 3 merges
  • In apps29Hugging Face Spaces

Top Spaces running it: webml-community/muse-glimmer-webgpu-kernels ♥ 17 · ukung/Muse-Glimmer-30B-DFlash-Abliterated ♥ 13 · akhaliq/Muse-Glimmer-30B ♥ 5 · apathy-exe/Muse-Glimmer-30B ♥ 4

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 64 GB Mac (unified memory) at INT4

About 18 GB at INT4, 36 GB at FP8, 72 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 18 GB of weights (estimate; KV cache and activations extra)INT418 GBFP8: about 36 GB of weights (estimate; KV cache and activations extra)FP836 GBBF16: about 72 GB of weights (estimate; KV cache and activations extra)BF1672 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 18 GB of weights (estimate; KV cache and activations extra)INT418 GBFP8: about 36 GB of weights (estimate; KV cache and activations extra)FP836 GBBF16: about 72 GB of weights (estimate; KV cache and activations extra)BF1672 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-08-09), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

SepOct2026-10-22: 1 repo commit2026-10-20: 2 repo commits2026-08-11: 2 repo commits2026-08-10: 5 repo commits2026-08-09: 6 repo commitsMuse Glimmer: 30B-parameter model optimized for always-on local agent workflows · 2026-08-10Muse Glimmer is a memory hierarchy disguised as a 30B Transformer · 2026-08-18Muse Glimmer · 2026-08-11Meta releases Muse Glimmer, 30B parameter open-weight edge model · 2026-08-10Muse Glimmer: Meta's open model built for always-on local agents. 30B parameters · 2026-08-10Muse Glimmer (High) Intelligence, Performance and Price Analysis · 2026-08-10Muse Glimmer: Benchmarks and analysis | Artificial Analysis · 2026-08-10Proving Muse Glimmer in Zero-Knowledge at 85 Tok/S on an H100 · 2026-08-11Experience of Running Muse Glimmer on Mac · 2026-08-11Fastest Inference Meta Muse Glimmer 30B on Apple · 2026-08-12Weights 2026-08-09 · 2026-08-09Weights 2026-08-09now · 3 providers
SepOct2026-10-22: 1 repo commit2026-10-20: 2 repo commits2026-08-11: 2 repo commits2026-08-10: 5 repo commits2026-08-09: 6 repo commitsMuse Glimmer: 30B-parameter model optimized for always-on local agent workflows · 2026-08-10Muse Glimmer is a memory hierarchy disguised as a 30B Transformer · 2026-08-18Muse Glimmer · 2026-08-11Meta releases Muse Glimmer, 30B parameter open-weight edge model · 2026-08-10Muse Glimmer: Meta's open model built for always-on local agents. 30B parameters · 2026-08-10Muse Glimmer (High) Intelligence, Performance and Price Analysis · 2026-08-10Muse Glimmer: Benchmarks and analysis | Artificial Analysis · 2026-08-10Proving Muse Glimmer in Zero-Knowledge at 85 Tok/S on an H100 · 2026-08-11Experience of Running Muse Glimmer on Mac · 2026-08-11Fastest Inference Meta Muse Glimmer 30B on Apple · 2026-08-12Weights 2026-08-09 · 2026-08-09Weights 2026-08-09now · 3 providers

What people built on it

263 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

263 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

191 quantized (Hugging Face model tree)56 fine-tunes (Hugging Face model tree)3 merges (Hugging Face model tree)13 adapters (Hugging Face model tree)191 quantized56 fine-tunes3 merges13 adapters
191 quantized (Hugging Face model tree)56 fine-tunes (Hugging Face model tree)3 merges (Hugging Face model tree)13 adapters (Hugging Face model tree)191 quantized56 fine-tunes3 merges13 adapters
GGUF unsloth/Muse-Glimmer-30B-GGUF for llama.cpp, Ollama, LM Studio 1.9M downloads MLX mlx-community/Muse-Glimmer-30B-4bit for Apple silicon 55.2k downloads NVFP4 RadixArk/Muse-Glimmer-NVFP4 for vLLM, SGLang 238.7k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

32 rows

Benchmarks 32 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (32)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported AA-LCR · Read long documents80High Reasoningmodel author2026-10-07
author-reported AIME 2026 · Reason through hard problems94.7High Reasoningmodel author2026-10-07
author-reported Beam128K65.1High Reasoningmodel author2026-10-07
author-reported CharXiv · Extract text from documents78.8High Reasoningmodel author2026-10-07
author-reported DeepSearch QA74.6High Reasoningmodel author2026-10-07
author-reported GDPVal-AA v2953High Reasoningmodel author2026-10-07
author-reported GPQA Diamond · Reason through hard problems83.5High Reasoningmodel author2026-10-07
author-reported Gaia243.3High Reasoningmodel author2026-10-07
author-reported HLE Text (AA)22High Reasoningmodel author2026-10-07
author-reported HPCT52.3% percentHigh Reasoningmodel author2026-10-07
author-reported IFBench · Chat and follow instructions77High Reasoningmodel author2026-10-07
author-reported MBCT41.5% percentHigh Reasoningmodel author2026-10-07
author-reported MCP Atlas (Public)75.5High Reasoningmodel author2026-10-07
author-reported MMMU-Pro · Understand images and video74High Reasoningmodel author2026-10-07
author-reported OSWorld-Verified65.9High Reasoningmodel author2026-10-07
author-reported OmniDocBench · Extract text from documents75.8High Reasoningmodel author2026-10-07
author-reported SWE-bench Pro · Write code with agents51.2High Reasoningmodel author2026-10-07
author-reported SWE-bench Verified · Write code with agents76High Reasoningmodel author2026-10-07
author-reported SciCode43.6High Reasoningmodel author2026-10-07
author-reported ScreenSpot Pro75.4High Reasoningmodel author2026-10-07
author-reported SkillsBench (with skills)44.3High Reasoningmodel author2026-10-07
author-reported Terminal-Bench · Write code with agents51.7High Reasoningmodel author2026-10-07
author-reported VCT37.0% percentHigh Reasoningmodel author2026-10-07
author-reported WildClawBench47.6High Reasoningmodel author2026-10-07
author-reported 𝛕3-Banking23.5High Reasoningmodel author2026-10-07
Show all 32 rows (7 more)
BenchmarkValueSetting / notesReported byDate
author-reported mmi-value (facebook/mmi)11.58—MMI evaluation2026-10-20
author-reported modality-presence-score (facebook/mmi)15.54—MMI evaluation2026-10-20
author-reported error-rate (facebook/mmi)41.1—MMI evaluation2026-10-20
author-reported pass-rate-strict (facebook/mmi)9.18—MMI evaluation2026-10-20
author-reported pass-rate-lenient (facebook/mmi)10.86—MMI evaluation2026-10-20
author-reported recall (facebook/mmi)14.28—MMI evaluation2026-10-20
author-reported precision (facebook/mmi)99.78—MMI evaluation2026-10-20

What the card says

1 cited sentence

Summary generated, cited

The architecture is a dense causal transformer with a dedicated perception encoder.[1]

  1. “Dense Causal Transformer with Perception Encoder”

Generated 2026-10-07 15:01 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

8 use cases

What it is good at 8 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

21 fields

Listed in LLMs

Facts

Pipeline tagimage-text-to-text
Librarytransformers
ArchitecturesMuseGlimmerForConditionalGeneration
Model typemuse_glimmer
Dtypebfloat16
Params by dtypeBF16: 29,776,626,688
Size bandL
MoEno
Context length131,072 (config.json max_position_embeddings)
PapersarXiv:2504.13181, arXiv:2602.06036, arXiv:2610.24628
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · MuseGlimmerForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · MuseGlimmerForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Muse-Glimmer-30B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)72 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)36 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)18 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Original model Family: meta-models/Muse-Glimmer 263 derivatives on the Hub

quantized: 191finetune: 56merge: 3adapter: 13

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 unsloth 575 1.9M
meta-models/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 meta-models 383 1.4M
RadixArk/Muse-Glimmer-NVFP4 quantized NVFP4 18.1B L — RadixArk 10 238.7k
RedHatAI/Muse-Glimmer-30B-FP8-block quantized -FP8 29.8B L apache-2.0 RedHatAI 13 148.1k
bartowski/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 bartowski 21 99.8k
unsloth/Muse-Glimmer-30B finetune — 29.8B L apache-2.0 unsloth 15 97.5k
0bserverx/Muse-Glimmer-30B-Heretic-Uncensored-GGUF quantized GGUF — — apache-2.0 0bserverx 28 94.8k
mlx-community/Muse-Glimmer-30B-4bit quantized -4bit 29.8B L apache-2.0 mlx-community 18 55.2k
NANI-Nithin/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 NANI-Nithin 6 34.7k
nvidia/Muse-Glimmer-30B-NVFP4 quantized NVFP4 19.8B L apache-2.0 nvidia 25 32k
unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit quantized -bnb 29.8B L apache-2.0 unsloth 14 29.2k
meta-models/Muse-Glimmer-30B-ExecuTorch-PTE quantized — — — apache-2.0 meta-models 42 23.8k
vmlinux/Muse-Glimmer-30B-ROCmFPX-GGUF quantized GGUF — — apache-2.0 vmlinux 19 22k
AtomicChat/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 AtomicChat 9 21.3k
RedHatAI/Muse-Glimmer-30B-NVFP4 quantized NVFP4 18.8B L apache-2.0 RedHatAI 18 16.8k
RedHatAI/Muse-Glimmer-30B-INT4 quantized INT4 7.95B M apache-2.0 RedHatAI 8 15.2k
kingjones777/Muse-Glimmer-30B-ROCmFP4-Strix-Halo-DFlash-GGUF quantized GGUF — — other kingjones777 4 13.6k
RadixArk/Muse-Glimmer-q4-MLX quantized MLX 27.9B L — RadixArk 12 13.1k
cvgro/Muse-Glimmer-30B-Abliterated-GGUF quantized GGUF — — apache-2.0 cvgro 3 12.8k
AaryanK/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 AaryanK 22 11.8k
mlx-community/Muse-Glimmer-30B-bf16 finetune -bf16 29.8B L apache-2.0 mlx-community 8 12k
cyankiwi/Muse-Glimmer-30B-AWQ-INT4 quantized AWQ 29.8B L apache-2.0 cyankiwi 11 10.5k
dealignai/Muse-Glimmer-30B-CRACK-GGUF quantized GGUF — — apache-2.0 dealignai 9 10.4k
z-lab/Muse-Glimmer-30B-DFlash2 finetune — 2.77B M apache-2.0 z-lab 15 9k
DevQuasar/meta-models.Muse-Glimmer-30B-GGUF quantized GGUF — — — DevQuasar 1 9.1k
mlx-community/Muse-Glimmer-30B-8bit quantized -8bit 29.8B L apache-2.0 mlx-community 4 7.9k
z-lab/Muse-Glimmer-30B-DFlash2-GGUF quantized GGUF — — apache-2.0 z-lab 12 7.5k
Abiray/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 Abiray 11 7.1k
bluevoid-pl/Muse-Glimmer-30B-pruned-GGUF quantized GGUF — — apache-2.0 bluevoid-pl 0 7k
incoai/Muse-Glimmer-30B-DFlash2-GGUF quantized GGUF — — apache-2.0 incoai 18 5.8k
mradermacher/Muse-Glimmer-30B-i1-GGUF quantized GGUF — — apache-2.0 mradermacher 1 6.5k
mlx-community/Muse-Glimmer-30B-OptiQ-4bit quantized -4bit 27.9B L apache-2.0 mlx-community 8 5.8k
OpenVINO/Muse-Glimmer-30B-int4-ov finetune int4 — — apache-2.0 OpenVINO 2 5.5k
mlasli/Muse-Glimmer-30B-Heretic-Abliterated-Q8_0-GGUF quantized Q8_0 — — apache-2.0 mlasli 1 5.5k
incoai/Muse-Glimmer-30B-DFlash2 finetune — 2.77B M apache-2.0 incoai 15 4.2k
mlasli/Muse-Glimmer-30B-Heretic-Abliterated-Q6_K-GGUF quantized Q6_K — — apache-2.0 mlasli 1 4.2k
prithivMLmods/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 prithivMLmods 2 4k
kingjones777/Muse-Glimmer-30B-Uncensored-ROCmFP4-GGUF quantized GGUF — — apache-2.0 kingjones777 0 4k
Distillio/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 Distillio 0 3.8k
ggml-org/Muse-Glimmer-30B-GGUF quantized GGUF — — apache-2.0 ggml-org 6 3.3k

Repository changelog

16 commits
changed in the last 7 daysRank band moves: 2026-10-23Repo updated: 2026-10-23Repo commits: 2026-10-22, 2026-10-20, 2026-10-20Other: 2026-10-24, 2026-10-24all changes

Repository changelog 16 commits

DateCommitSHA
2026-10-22 Create .eval_results folder and add mmi results (#68) cec94fc
2026-10-20 Update README.md 7af5801
2026-10-20 Update README.md aae9566
2026-08-11 Fixing duplicate reasoning effort - formatting (#35) a4e59da
2026-08-11 Category name change 97c77df
2026-08-10 Update generation_config.json (#31) 90625aa
2026-08-10 Add Apache-2.0 license and usage policy f84ecc3
2026-08-10 Add transformers as library_name (#2) e13d2ae
2026-08-10 Muse Glimmer 30B (#1) 46ac57a
2026-08-10 Update README.md 23e858a
2026-08-09 Update README.md 049fe29
2026-08-09 Update README.md bb2d63d
2026-08-09 Update README.md 1eec7a7
2026-08-09 Update README.md d4e5edb
2026-08-09 first pass at the model card. 5f5a77e
2026-08-09 initial commit 45ad554

All providers

3 providers · 3 endpoints

Where to run 3 providers · 3 endpoints

Hosted endpoints observed for Muse-Glimmer-30B, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.30/M.

ProviderModel refContextQuant$/M in$/M outtok/sTTFTToolsStatusObserved
DeepInfra
hf, deepinfra, openrouter
 meta-models/Muse-Glimmer-30B 128K bf16 $0.30 $1.20 97.5 751 ms yes live 2026-10-25
Together AI
hf, openrouter
 meta-models/Muse-Glimmer-30B 128K unknown $0.35 $1.50 82.8 268 ms yes live 2026-10-25
Featherless
hf
 meta-models/Muse-Glimmer-30B — — — — — — — live 2026-10-25

Articles and coverage

24 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

Muse Glimmer: Meta's open model built for always-on local agents. 30B parameters
developer.meta.com2026-08-10about the Muse Glimmer familyHN 12 pts · 2 comments285 words
Muse Glimmer is a 30B parameter model built for agents that don't stop with reliable tool-calling and persistent state across restarts. It enables local deployments on a single consumer GPU or Mac while supporting self-managed memory across hours-long sessions. Engineers can access the full model card on Hugging Face for detailed specifications.written, reviewed, cited

Analysis8

Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
research.meta.ai2026-08-10HN front pageabout the Muse Glimmer familyHN 1,209 pts · 638 comments1,186 words
Meta releases Muse Glimmer, a 30-billion-parameter open model licensed under Apache 2.0 for local agentic workflows. The model uses 4-bit quantization to shrink weights to under 20 GB, fitting within a 24 GB or 32 GB envelope for consumer GPUs. A lightweight DFlash drafter enables faster generation, with speed validated on MacBook M4-Max, M5-Max, and RTX-5090 hardware.written, reviewed, cited
  • Muse Glimmer trains on Muse Spark's outputs using logit distillation while leveraging a similar data mix as the teacher model.
  • The model accepts interleaved text and images through a dedicated perception encoder to interpret screenshots, charts, and documents.
  • Muse Glimmer supports different reasoning strengths to select the right balance between quality and speed for controllable effort.
Muse Glimmer is a memory hierarchy disguised as a 30B Transformer
abstractextraordinary.com2026-08-18about the Muse Glimmer familyHN 16 pts · 9 comments4,110 words
Meta's Muse Glimmer is a 30B-parameter multimodal agent designed to run autonomously on consumer hardware without cloud infrastructure. The architecture splits the 52 text Transformer blocks into 39 sliding-attention layers with a 2,048-token window and 13 full-attention layers that operate without RoPE to enable content-based retrieval over the full 131,072-token context. Four-bit quantization reduces the 55.46 GiB BF16 checkpoint to fit within 24 or 32 GB envelopes, resulting in a total active KV cache of approximately 1.70 GiB per sequence.written, reviewed, cited
  • The 13 full-attention layers still perform quadratic attention work during prefill, meaning a 131K prefill is not equivalent to a 4K prefill.
  • Global layers operate on position-aware representations because every key and value has already passed through three RoPE-equipped local layers.
Muse Glimmer
simonwillison.net2026-08-11authorityabout the Muse Glimmer familyHN 4 pts · 1 comments598 words
Meta released Muse Glimmer, a 30B parameter vision model under a clean Apache 2.0 license optimized for local deployment. The model achieves strong success rates on benchmarks including DeepSearch QA, MCP-Atlas, and SWE-Bench while handling extended workflows. An 18.16 GB version runs efficiently on machines with 32 GB of RAM or more, leaving space for other applications.written, reviewed, cited
Muse Glimmer (High) Intelligence, Performance and Price Analysis
artificialanalysis.ai2026-08-10authorityabout the Muse Glimmer familyHN 2 pts · 0 comments1,034 words
Muse Glimmer: Benchmarks and analysis | Artificial Analysis
artificialanalysis.ai2026-08-10authorityabout the Muse Glimmer family940 words
Proving Muse Glimmer in Zero-Knowledge at 85 Tok/S on an H100
attestable.com2026-08-11about the Muse Glimmer familyHN 4 pts · 0 comments806 words
Experience of Running Muse Glimmer on Mac
holaclaw.ai2026-08-11about the Muse Glimmer familyHN 4 pts · 0 comments1,468 words
Fastest Inference Meta Muse Glimmer 30B on Apple
basecompute.co2026-08-12HN 2 pts · 1 comments159 words

Community1

Meta releases Muse Glimmer, 30B parameter open-weight edge model
twitter.com2026-08-10about the Muse Glimmer familyHN 18 pts · 1 comments
All coverage tracked for this model (24) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Muse Glimmer 30B Architecture Notes | Sebastian Raschka, PhDsebastianraschka.com—2026-10-07exa
Muse Glimmer 30B: An Architecture and Hands-On Look at Meta's New Mid-Range Model - DEV Communitydev.to2026-09-182026-10-07exa
Muse Glimmer is a memory hierarchy disguised as a 30B Transformerabstractextraordinary.com2026-08-182026-10-07hn
Meta Muse Glimmer 30B: Full Review (September 2026) - AIToolsReviewaitoolsreview.co.uk2026-08-172026-10-07exa
Muse Glimmer 30B: A Comprehensive Technical Analysis - Local AI Zonelocal-ai-zone.github.io2026-08-152026-10-07exa
Qwen 3.8 27B vs. Nemotron 3.5 vs. Muse Glimmer on an RTX 5090kgptalkie.com2026-08-142026-10-07hn
Fast, on Device Agentic AI with Muse Glimmer on ExecuTorchpytorch.org2026-08-142026-10-07hn
Run OpenClaw with Muse Glimmer Locally on Macholaclaw.ai2026-08-122026-10-07hn
Fastest Inference Meta Muse Glimmer 30B on Applebasecompute.co2026-08-122026-10-07hn
Muse Glimmer 30B: 24/32 GB, DFlash, benchmarks vs Qwen3.6-27B — AI on Macai-on-mac.com2026-08-122026-10-07exa
Proving Muse Glimmer in Zero-Knowledge at 85 Tok/S on an H100attestable.com2026-08-112026-10-07hn
Experience of Running Muse Glimmer on Macholaclaw.ai2026-08-112026-10-07hn
Muse Glimmersimonwillison.net2026-08-112026-10-07hn
Muse Glimmer: Meta’s 30B Model Built for Efficient Inferencekaitchup.substack.com2026-08-112026-10-07exa
Zuck rekindles open weights Llama drama with Muse Glimmertheregister.com2026-08-102026-10-07hn
Muse Glimmer (High) Intelligence, Performance and Price Analysisartificialanalysis.ai2026-08-102026-10-07hn
Meta open-sources Muse Glimmer: 30B agent model runs locally under 20GB VRAMainexusdaily.vercel.app2026-08-102026-10-07hn
Muse Glimmer from Meta Superintelligence Labs is now availableollama.com2026-08-102026-10-07hn
Meta releases Muse Glimmer, 30B parameter open-weight edge modeltwitter.com2026-08-102026-10-07hn
Muse Glimmer: Meta's open model built for always-on local agents. 30B parametersdeveloper.meta.com2026-08-102026-10-07hn
Muse Glimmer: 30B-parameter model optimized for always-on local agent workflowsresearch.meta.ai2026-08-102026-10-07hn, exa
Muse Glimmer: Benchmarks and analysis | Artificial Analysisartificialanalysis.ai2026-08-102026-10-07exa
Muse Glimmer 30B Hardware and Benchmark Guide | Wavectwavect.io2026-08-102026-10-07exa
Muse Glimmer 30B Review — Benchmarks, Pricing & AI Panel Verdict | TopReviewed.aitopreviewed.ai2026-08-092026-10-07exa

Papers

3 papers

Papers 3

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • Qwen3.8-27B · 27.8B
    same size band (L), same task (image-text-to-text), more providers, permissive license
  • gemma-4-31B-it · 31.3B
    same size band (L), same task (image-text-to-text), more providers, permissive license
  • Qwen3.6-27B · 27.8B
    same size band (L), same task (image-text-to-text), permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:34 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:01 UTC · Last verified 2026-10-25

Follow Muse-Glimmer-30B

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.