Muse Glimmer [What It Changes]

TL;DR

• Muse Glimmer is Meta Superintelligence Labs' 30-billion-parameter open-weight agent model, released 10 August 2026 under Apache 2.0, meant to run on one consumer GPU after 4-bit quantization.

• Meta's own table puts MCP Atlas at 75.5; that is a vendor score, not a third-party ranking.

• Full precision needs more than 55 GB; the 4-bit build is under 20 GB. Weights are on Hugging Face. This is a downloadable model, not a hosted Meta product with an uptime SLA.

• Day-zero runtimes include llama.cpp, MLX, ExecuTorch, vLLM, SGLang, Ollama, and LM Studio; hosted options include Together, Fireworks, and OpenRouter.

What Muse Glimmer is

Muse Glimmer is a local agent model: a dense transformer plus a perception encoder, distilled from a larger Muse teacher, licensed so a shop can download weights and run them on a machine it owns.

A two-truck HVAC shop, a ten-person marketing agency, and a solo clinic should care because the alternative is sending every work order, client deck, and chart into a hosted API. If the model fits under 20 GB at 4-bit, the constraint that used to be "we cannot afford a private GPU cluster" becomes "we have one 24 GB card, or we do not." That is a buying question, not a research paper.

This hub is current as of 10 August 2026, the date on Meta's research post and AI Weekly. MarkTechPost's source-pack URL returned 403 when fetched; it is not cited.

Why a local 30B agent is the story

Most agent demos still assume the cloud. Muse Glimmer is Meta's bet that a 30B-class model, trained for tool calls, failure recovery, and screenshots, can sit next to the files. The Hugging Face model card calls out coding agents, local assistants, and LLM-as-a-judge as intended uses, and says audio is out of scope.

The Apache 2.0 license is a royalty-free copyright and patent grant with an AS-IS disclaimer. It is not a support contract.

What Meta published

According to Meta AI Research, Muse Glimmer is a 30-billion-parameter model under Apache 2.0, small enough for a Mac or PC with a single consumer GPU. According to the same research post, full precision needs over 55 GB and 4-bit quantization shrinks the language model to under 20 GB.

According to AI Weekly, 4-bit is the consumer-GPU path and Meta's tables put Glimmer ahead of Gemma4-31B and Qwen3.6-27B on MCP Atlas, GAIA2, SWE-Bench Pro, and AIME 2026. AI Weekly flags those as self-reported.

According to the Hugging Face model card, ~29.6B is the parameter count, 131,072+ is the context length, and 75.5 is Muse Glimmer's MCP Atlas (Public) high-reasoning score versus 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B.

The Meta developer page repeats the same MCP Atlas 75.5 row and points to developer docs. According to the Hugging Face blog, the architecture is a 2B ViT-style perception encoder plus a 28B text decoder.

Architecture, without equations

The model card lists a dense causal transformer with a perception encoder: 52 layers, hidden size 6656, GQA 32 query / 2 KV heads, SwiGLU FFN, sliding window 2048, RoPE on local layers, vocabulary 202,048, max 4,096 visual tokens per image, input text plus image, output text. Knowledge cutoff: 4 January 2026. Training data includes public sources, third-party data, and Meta products, plus vendor networks.

The perception encoder is ~1.8B ViT-G/14 (model card) / ~2B (Hugging Face blog). Hugging Face describes hybrid attention as three sliding-window layers then one full-attention NoPE layer, repeated to 52 layers.

Speculative decoding uses a DFlash drafter that proposes 16-token blocks. Meta reports decode speedups on the K-Quant-17GB build with that drafter.

Benchmarks are Meta's table

Sources: Hugging Face model card; Meta developer page; methodology report. Vendor table, not a league.

Qwen wins some rows (SWE-Bench Verified 77.2, TerminalBench 60.7, OSWorld 75.6). Gemma wins GPQA Diamond 85.7 and has a lower AgentDojo attack-success rate. Do not quote MCP Atlas 75.5 as "the best 30B model."

Memory and speed

Sources: model card; research post. Batch size 1, greedy decoding; M4/M5 via ExecuTorch, RTX via llama.cpp.

AI Weekly independently restates the 3.1x RTX 5090 and 1.8x M5 Max speedups as Meta-listed 4-bit figures.

The Hub card showed 609,919 downloads in the last month at fetch time, 12 adapters, 38 finetunes, and 160 quantized variants. Those are Hub counters, not quality scores.

How to run it

Weights: huggingface.co/meta-models/Muse-Glimmer-30B. Meta lists Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, SGLang, Together, Fireworks, and OpenRouter. Hugging Face shipped day-0 Transformers, llama.cpp, and vLLM-with-transformers-backend support, plus a GGUF repo.

Sampling on the model card: temperature 1.0, top_p 0.95, top_k 64. Reasoning strength is a system-prompt knob: low / medium / high / xhigh.

The research post says Muse Glimmer was assessed under Meta's Advanced AI Scaling Framework. The card says it is not "Frontier AI" under that framework because it is generally less capable than Muse Spark; chem/bio, cyber, and loss-of-control were rated moderate or lower.

Intended-use limits: not for under-18 downloaders; deployers must mitigate if a system may be used by minors; not for illegal use; video is processed as frames; quantized inference may differ on edge cases.

USTA analysis

USTA analysis. Inputs from Meta's memory table: full precision >55 GB; 4-bit language model <20 GB; K-Quant-17GB targets 24 GB VRAM with 1.0% average degradation; RTX 5090 decode 74.9 tok/s baseline vs 233.4 tok/s with DFlash (3.1x).

• 20 / 55 ≈ 0.36, so the 4-bit language-model envelope is about 36% of the >55 GB full-precision figure, a reduction of about 64% of that memory number.

• 233.4 / 74.9 ≈ 3.12, which matches Meta's 3.1x rounding on the RTX 5090 row.

• 1.0% degradation on the 17 GB quant is Meta's average across 15 benchmarks, not a guarantee on your invoice parser.

No price is derived. The analysis only divides memory and throughput figures Meta printed.

What a small shop can actually do with it

If you already extract invoices and work orders, a local 30B agent is a candidate for "read the PDF on this PC, draft the coding, do not send." Accounting teams comparing practice management software or Drake versus ProConnect should keep tax engines as the system of record; Muse Glimmer can draft, not file. Reporting stacks in Fathom versus Jirav stay on the ledger.

Form-to-CRM automation and executive-assistant tasks are the right shape: one GPU, one folder, one review queue. The state of small-business automation still starts with the queue, not the model card.

Teams already routing documents through US Tech Automations can treat Muse Glimmer as a model swap on the extraction box, not a rebuild of the workflow. Keep send, file, and post behind a person. If the shop has no 24 GB GPU, use a hosted endpoint from Meta's partner list and keep the same review step.

A shop-floor path from download to draft

Do not start with the full-precision 55 GB build. Start with the K-Quant-17GB weights on a 24 GB card, or with a listed host if you have no such card. The Ollama library page, LM Studio model page, Unsloth docs, Together model page, and Meta's partner list for Fireworks and OpenRouter are distribution doors, not quality rankings.

Pin the Hugging Face revision. The usage policy on the Hub is the acceptable-use file; Apache 2.0 does not replace it. The Perception Encoder paper is the vision-backbone citation if you need to tell a security reviewer what the image tower is.

For accounting work, keep the tax engine as the filer. Drake versus ProConnect and practice-management software are still the systems that talk to the IRS. Muse Glimmer can read a scanned organizer and draft a checklist. It cannot sign.

Run one job: a folder of invoices, a prompt that says "extract vendor, date, amount, do not email anyone," and a human who compares three rows to the PDF. If that fails, stop. If it holds, add a screenshot of the vendor portal as a second input and see whether the perception encoder helps or hallucinates a button.

Log every tool call. Meta's card says failure recovery is trained in; your shop still needs a transcript. Put that transcript on the same path as form-to-CRM so a reviewer can see the source image next to the draft.

Quantized GGUF builds live in the Hub collection Meta and Hugging Face describe in the blog. A community quant that is not in Meta's table is a different product. Do not mix its speed claims with the 3.1x RTX 5090 row.

Honest limits

This is not a Meta SLA. Apache 2.0 is AS-IS. Benchmarks are Meta-chosen. Safety rows show Glimmer is not the lowest attack-success model in the table. The card recommends extra guardrails and human confirmation for irreversible actions. Hugging Face's blog is enthusiastic packaging, not an independent bake-off.

AI Weekly's 403-free alert is a secondary digest of Meta and Hugging Face; it is not a lab result.

Signal vs Speculation

Signal (sourced). 10 August 2026 open-weight release; Apache 2.0; ~29.6B; 131K context; 4-bit under 20 GB; Meta MCP Atlas 75.5; DFlash 3.1x on RTX 5090; Hub weights and day-0 runtime list; not a hosted Meta product.

Speculation (our read). If the 24 GB path holds up outside Meta's 15-benchmark average, privacy-sensitive SMBs will try local draft agents this year. Our read: 12–36 months of "one GPU in the back office" is more likely than "replace the hosted copilot," because tool-calling still needs connectors, logging, and a person on send. Our read is not a claim that 75.5 MCP Atlas will replicate on your MCP server.

According to AICPA, 62% of firms reported cloud-workflow adoption.

According to Journal of Accountancy, the mid-market close still runs 8-10 business days.

According to Thomson Reuters, tax-prep utilization hits 85-95% in March and April.

According to NFIB, 44% of small businesses cite time-management. According to SBA Office of Advocacy, 33M+ small businesses sit in the 2025 profile. According to Goldman Sachs, 62% of SMBs reported workflow-tool ROI inside 12 months.

Key Takeaways

• Muse Glimmer is a downloadable 30B local agent model, not a Meta cloud SKU.

• Treat 75.5 MCP Atlas as Meta's number.

• Budget a 24 GB card for the 17 GB quant plus KV cache and encoder, or use a listed host.

• License is Apache 2.0 AS-IS; add your own confirmation on writes.

• Plug it in as a model swap on an existing review workflow.

FAQ

Is Muse Glimmer free to use commercially?

Direct-answer first sentence: the weights are under Apache 2.0, which grants a royalty-free copyright and patent license for use and distribution, with an AS-IS disclaimer. Read the license and Meta's usage policy on the Hub; "open weights" is not "Meta will support your clinic."

Will it run on a laptop with 16 GB of unified memory?

Direct-answer first sentence: Meta's 4-bit language-model figure is under 20 GB, and the 17 GB quant targets 24 GB, so 16 GB is below those envelopes. Use a smaller quant only if a third party publishes one and you accept the extra quality risk; Meta's own table does not list a 16 GB target.

Is 75.5 MCP Atlas a third-party ranking?

Direct-answer first sentence: no; it is Meta's published score on a vendor table. Hugging Face reprints it and says scores are as published.

Does Meta host Muse Glimmer with an SLA?

Direct-answer first sentence: no; the research post and this hub treat it as downloadable weights. Together, Fireworks, and OpenRouter are listed as partners, each with their own terms.

Can it see screenshots of a dispatch board?

Direct-answer first sentence: yes, the perception encoder accepts images, including screenshots, alongside text. Video is frames only; audio is unsupported.

What to do next

If you have a 24 GB GPU and a folder of documents that must not leave the building, download the K-Quant-17GB build and run one draft-only job with a human review. If you do not, skip the hero demo. Route the same draft-and-review path through agentic workflows so US Tech Automations keeps the exception log when the local model is swapped or turned off.