Meta put a 30B agent model on a laptop — Apache 2.0, no server required
Meta released Muse Glimmer on August 10, a 30-billion-parameter agentic model distilled from Muse Spark 1.2 and published under Apache 2.0, compressed to roughly 4-bit so it runs on a single consumer GPU or a Mac with no network call
The three lines
- Muse Glimmer 30B — distilled from Muse Spark 1.2, weights released under Apache 2.0
- Over 55GB at full precision, compressed to roughly 4-bit and paired with block-level speculative decoding for agent-loop speed
- Leads its size class on MCP-Atlas (75.5) and DeepSearch QA (74.6); trails on computer-use and terminal work
Key questions
- What is Muse Glimmer
- A 30-billion-parameter open-weights model Meta released on August 10, distilled from its larger Muse Spark 1.2. It targets agentic work — scheduling, document organization, multi-step tasks a system carries out on its own rather than answering once. The point is the footprint: it runs on a single consumer graphics card or a Mac, not in a datacenter.
- Apache 2.0 — is it really open source
- The license on the weights is genuinely permissive: commercial use, modification and redistribution, without user-count conditions. But 'open source' deserves care as a term. What was published is the trained weights. The training data and training code were not.
- Why does running locally matter
- Until now, capable agents mostly ran on someone else's servers. Documents you handed them left the building, an outage stopped them, and every call cost money. A 30B model running locally changes all three. It also relocates the risk: on a prompt-injection benchmark this model showed a 28.4% attack success rate, meaning the defense problem moves onto the user's device.
Last week the question in AI was how to restrain the strongest model — OpenAI had paused one of its own. On August 10 Meta answered a different question: how far can the model on your own laptop go?
Muse Glimmer is a 30-billion-parameter model distilled from the larger Muse Spark 1.2, with weights released under Apache 2.0. Mark Zuckerberg framed the release around the idea that superintelligence should reach everyone rather than a handful of firms.
1. How 30B fits on consumer hardware
At full precision a 30-billion-parameter model needs more than 55GB of memory. That does not fit on personal hardware. Meta did two things.
First, compression to roughly 4 bits per parameter. Cutting the bits used to store each weight cuts the memory requirement proportionally. Second, block-level speculative decoding — predicting several tokens at once and rolling back only the wrong ones, which raises throughput on the same silicon.
The second is not a nicety. An agent does not answer once; it calls a tool, reads the result, decides again, dozens of times per task. If a single response is slow, the whole loop becomes unusable. Fitting inside an agent loop was the real design constraint, not fitting in memory.
2. Where it wins and where it doesn't
| Benchmark | Muse Glimmer 30B | What it measures |
|---|---|---|
| MCP-Atlas | 75.5 (vs 54.2 / 62.5) | Tool connection and orchestration |
| DeepSearch QA | 74.6 | Multi-step research |
| Gaia2 | 43.3 | Composite real-world tasks |
| SWE-Bench Pro | 51.2 | Code repair and debugging |
| Siren AgentDojo — attack success | 28.4% (Gemma 25.6 / Qwen 40.3) | Prompt-injection resistance — lower is safer |
| Siren AgentDojo — utility | 94.2 (highest of three) | Staying useful while defending |
The top four rows say the model is at or near the front of its size class at planning multi-step work and calling tools. Meta itself notes it trails on computer-use and terminal tasks.
The bottom two rows are the more interesting pair. Prompt injection — hiding instructions inside a document or web page so the model obeys the attacker rather than the user — succeeded 28.4% of the time here, below Qwen (40.3%) and above Gemma (25.6%). Utility was the highest of the three at 94.2, which supports Meta's claim that safety was not bought with capability. It does not change what 28.4% means: more than one attempt in four gets through.
3. What moves onto the device
Data. Documents no longer have to leave the building to be processed. That opens agentic work on material whose export is the problem — internal records, contracts, medical files.
Cost. Per-call billing disappears. Always-on agents were structurally incompatible with metered APIs; that constraint is gone.
Responsibility. Here the direction reverses. When the model sat on a vendor's servers, injection defense was the vendor's job. On your laptop it is yours. That is where the 28.4% starts to matter in practice.
Meta says a more capable Muse Spark release will follow. The weights are on Hugging Face, and AMD published a same-day guide for running the model on Ryzen AI Max and Radeon GPUs.
4. What remains unverified
Every benchmark cited is Meta's own measurement. There is no independent reproduction yet, and the version and configuration of the comparison models have not been externally audited. The quality cost of 4-bit compression was not disclosed — and whether the benchmark scores were produced before or after compression materially changes how to read them. Training-data composition and the size of the parent Muse Spark 1.2 remain undisclosed.
On terminology, this publication's standard: what was released is weights, not training data or code. The grading of openness is covered in "What is an open-source AI model", and the mechanics of running models locally in today's companion piece, "What is on-device AI".
Sources
- Meta AI Research — Introducing Muse Glimmer
- VentureBeat — Meta returns to open source with an Apache 2.0 30B agent model
- CNBC — Meta launches Muse Glimmer open-weight AI model
- Engadget — Muse Glimmer can run on a single computer
- Phoronix — Meta publishes Muse Glimmer as 30B open agentic model
- Hugging Face — meta-models/Muse-Glimmer-30B