Meta Superintelligence Labs published Muse Glimmer on August 10 under Apache 2.0. It is 30 billion parameters, dense rather than sparse, and it needs more than 55 GB of memory to run at full precision. It also ships with a quantized build that fits in 17 gigabytes and gives up about 1.0% of its average score across fifteen benchmarks.
Most outlets led with the licensing reversal, and it is a real reversal. The quantized build is the part that changes what a team can deploy, because a capable agent now fits on a graphics card they already own.
What Meta published
Glimmer is a dense causal transformer with a separate perception encoder, distilled from Meta’s closed flagship Muse Spark 1.2 using logit distillation. The model card lists roughly 29.6B parameters across 52 layers, grouped-query attention at 32 query heads to 2 KV heads, and an attention pattern of three local layers to one global layer with a 2,048-token sliding window. Context is 131,072 tokens, the vocabulary is 202,048, and the knowledge cutoff is January 4, 2026. The vision tower is a roughly 1.8B ViT-G/14 accepting up to 4,096 visual tokens per image. Text and images go in; only text comes out.
Three builds ship. Full-precision BF16 is the 55 GB one. K-Quant-Dynamic targets a 32 GB card and gives up 0.2% on average. K-Quant-17GB targets a 24 GB card and gives up 1.0%, measured by Meta across fifteen common evaluations.
Speed comes from DFlash, a block-diffusion drafter that predicts 16 tokens in a single forward pass and runs five layers of its own. On an RTX 5090 at batch size one with greedy decoding, Meta reports decode throughput going from 74.9 to 233.4 tokens per second, a 3.1x improvement. On an M5 Max the same technique takes 26.6 to 50.2 tokens per second, and on an M4 Max, 23.7 to 37.8.
Weights are on Hugging Face. Runtime support landed immediately across vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, and ExecuTorch for mobile and edge, and Meta documents the model working inside OpenClaw and similar agent orchestration frameworks.
The opposite bet from the trillion-parameter race
Two weeks ago Moonshot published Kimi K3’s weights on schedule, 1.56 terabytes of files that need roughly 1.4TB resident before you allocate a single byte of KV cache. Running them means an eight-way B300 node or two nodes of B200s with fast interconnect. Last week Alibaba shipped Qwen3.8-Max, a 2.4-trillion-parameter mixture of experts activating 95 billion per token, and committed to open-weighting it during the week of August 10 alongside a smaller Qwen3.8-27B.
Those releases work as transparency guarantees. They tell you what is inside a frontier system, and they hedge against a vendor raising prices or pulling a model, both of which are worth real money. Neither release delivers access, though, since a 2.78-trillion-parameter K3 being downloadable does not put it on anyone’s desk.
Glimmer inverts that proposition. It is a smaller model and it sits well behind the frontier, a point Meta makes itself: the company rated it moderate or lower across chemical and biological, cyber, and loss-of-control categories under its Advanced AI Scaling Framework, and concluded it does not meet its own internal definition of a frontier system. In exchange you get a file set that runs on a consumer card without an account or a metered token bill. The timing sharpens the contrast, since Alibaba’s 27B was promised for this same week and was not on Hugging Face when Meta’s model landed.
Where the benchmarks land
Meta compares Glimmer against Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B, which is the right peer group for a model of this size.
| Benchmark | Muse Glimmer | Gemma4-31B | Qwen3.6-27B |
|---|---|---|---|
| MCP Atlas | 75.5 | 54.2 | 62.5 |
| DeepSearch QA | 74.6 | 61.7 | 71.1 |
| WildClawBench | 47.6 | 37.6 | 43.2 |
| GAIA2 | 43.3 | 36.4 | 40.0 |
| AA-LCR | 80.0 | 68.3 | — |
| SWE-Bench Pro | 51.2 | 36.9 | — |
| SWE-Bench Verified | 76.0 | — | 77.2 |
| TerminalBench 2.1 | 51.7 | — | 60.7 |
| OSWorld-Verified | 65.9 | — | 75.6 |
| AIME 2026 | 94.7 | — | — |
| GPQA Diamond | 83.5 | — | — |
The pattern is consistent enough to plan around. Glimmer wins decisively on tool-call orchestration and search-shaped agent work, taking MCP Atlas by 13 points over Qwen and 21 over Gemma. It loses on sustained terminal sessions and computer control, trailing Qwen3.6-27B by nine points on TerminalBench 2.1 and nearly ten on OSWorld-Verified. On repository-scale software engineering the two are effectively tied.
The model is good at deciding which tool to call and stitching multi-step research together, and merely competitive at driving a shell for an hour. If your workload is document triage, internal search, structured extraction, or LLM-as-judge evaluation, you are pointing it at the half of that table where it wins.
Four months from closed to Apache 2.0
In April, Meta launched Muse Spark as a closed model, the first release from the restructured Superintelligence Labs under Alexandr Wang, and broke with the open-weight tradition that Llama had built. Four months later the company is shipping under Apache 2.0, which is more permissive than any Llama license Meta ever wrote. There is no monthly-active-user threshold, no attribution requirement in your product interface, and no acceptable-use rider. Set against the bespoke commercial gates Moonshot attached to K3, Meta’s terms carry the fewest strings of anything currently shipping with open weights.
Zuckerberg wrapped it in a 6,500-word essay titled “The Future Is for Everyone,” arguing that the largest risk from advanced AI is concentrated control rather than the failure modes usually cited. Superintelligence should be distributed rather than centralized, he wrote, so that every person has “the ability to direct it.” Meta also says it will open the weights for Muse Spark 1.2 itself in the coming weeks.
Treat that promise the way we treated Moonshot’s and Alibaba’s, as a date rather than a delivery. Moonshot hit its commitment and earned the credit for it. Alibaba’s 27B has not appeared yet, and Meta’s flagship stays unreleased until the files exist.
Two caveats belong on the label. The Open Source Initiative’s position is that these releases are open-weight rather than open-source, since Meta publishes neither training data, training code, nor complete architecture documentation. The Register’s counterpoint is worth keeping alongside the announcement, roughly that distribution and monetization still route through Meta and a handful of cloud providers regardless of who can download the weights. Both objections are correct and neither changes what the file does on your hardware.
What it takes to run this yourself
The hardware floor is a 24 GB card or an Apple Silicon machine with enough unified memory. That is one desktop GPU or a well-specced laptop, not a procurement cycle. It is also the part Meta’s coding agent release last week could not offer, since the binary ran on your machine but the model behind it billed per token, with a cheaper tier that traded training rights on your code.
Read the throughput numbers before you plan around them. An RTX 5090 at 233 tokens per second runs a real agent loop. An M4 Max at 37.8 tokens per second is comfortable for chat and slow for an agent that reasons over long context and calls tools repeatedly, because every step in that loop pays the full decode cost. Budget the laptop for interactive use and a desktop card for anything meant to run unattended.
The gate that will stop most teams is security rather than speed. On the Siren AgentDojo adversarial benchmark, Meta reports Glimmer at a 28.4% attack success rate with 94.2 utility. Gemma4-31B sits at 25.6 and Qwen3.6-27B at 40.3, so Glimmer is mid-pack. Mid-pack still means roughly one in four prompt-injection attempts succeeds. An always-on local agent with filesystem and browser access, pointed at untrusted inbound email, is not a configuration that number supports. Self-hosting does not fix that, and it can make the blast radius worse, because a compromised agent holding local credentials reaches further than one sandboxed inside a vendor’s cloud.
The deployment hygiene is on you rather than on Meta: bind the inference server to loopback, require auth on the endpoint, scope tool permissions per workflow instead of granting the agent everything, and log tool calls at the application layer where you can audit them.
That is a longer list than “download and run,” and it is still shorter than what a hosted agent would ask you to accept. Meta kept its best work closed for four months, then published a 30B under Apache 2.0 and promised the flagship next. The weights are worth putting into a pilot this month on the workloads Glimmer measures well. Whether the position behind them holds is a question Spark 1.2 answers in a few weeks, and until those files exist there is nothing to evaluate but a press cycle.
Sources
- Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device — Meta AI Research
- meta-models/Muse-Glimmer-30B model card — Hugging Face
- Muse Glimmer model documentation — Meta for Developers
- Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU — MarkTechPost
- Meta’s ‘open source’ Muse Glimmer model can run on a single computer — Engadget
- Zuckerberg: AI’s biggest risk is one entity with too much control — Axios
- The future is for billionaires, the rest of us will get open weight AI models, maybe — The Register
- Meta Publishes Muse Glimmer As 30B Open Agentic Model — Phoronix