Meta has released Muse Glimmer, an open-weight language model of roughly 30 billion parameters built specifically to run AI agents on a machine you own rather than in a data center. The release is under Apache 2.0, one of the most permissive software licenses in use, and Meta's own sizing puts the quantized model under 20 GB -- small enough for a single high-end consumer graphics card. It is the first time Meta has aimed a purpose-built agent model at local hardware rather than at the top of a benchmark table.
Key facts
- About 29.6 billion total parameters, including the model's image-understanding encoder, in a dense transformer design.
- Compressed to roughly 4-bit precision, the language model fits under 20 GB, targeting a 24 GB or 32 GB memory envelope.
- Released August 2026 by Meta under Apache 2.0, with text and image input, text output, and no audio support.
- Primary source: Meta's release post and the Hugging Face model card.
The interesting part is not the parameter count. It is what Meta built around it. An "always-on local agent" -- something that watches your files, answers on your machine, holds a long conversation in memory, and calls tools without a round trip to a server -- has a very specific hardware problem. It has to be resident in memory all the time, which means it has to be small enough to leave room for everything else. Meta's description of the model is exactly that: it is, in the company's words, "optimized for always-on local agent workflows."
To hit that target, Meta did three things. First, it distilled the model from the larger Muse Spark line, meaning the smaller model was trained to imitate the bigger one's outputs rather than learning from raw text alone -- the same technique behind most small models that punch above their weight, explained further in our lesson on distillation. Second, it shipped a companion speed trick called DFlash: a small drafter network that guesses an entire block of sixteen words at a time, which the main model then checks in a single pass. That is speculative decoding, and it is the difference between a local model that feels responsive and one that feels like waiting.
Third, and most consequential for memory, Meta used a mixed attention pattern. Most layers only look at a sliding window of 2,048 recent words; every fourth layer looks at the whole conversation. The repeating pattern runs local, local, local, global, and it supports a context window past 131,072 tokens. Think of it like a team where most people only track the current meeting and one person keeps the whole project history. You get long memory without paying for long memory at every layer -- the same tradeoff we cover in sparse attention and context windows.
The practical result is the story. Full-precision weights would need more than 55 GB. Quantized to about 4 bits -- a compression technique described in our quantization lesson -- the released path lands under 20 GB, leaving room for the key-value cache, the image encoder, and the drafter. Independent testers on Reddit confirmed the fit quickly, reporting a working setup on a single RTX 3090 at 22 to 23 GB of memory use. Within hours the community had packaged it: the Unsloth GGUF mirror went up with instructions for llama.cpp, Ollama, vLLM, and SGLang. That kind of same-day packaging is the clearest signal that the local-model crowd treated this as a real deployable base rather than a demo.
The reception was large and split. The story hit number three on Hacker News with 1,045 points and 579 comments, and the Reddit launch thread drew 1,466 upvotes. Supporters reported stable day-one tool use and good behavior as a coordinator for multi-model setups. The pushback was specific: several testers said it falls behind Qwen 3.6 27B on coding, and one comparison found Glimmer at 8-bit precision worse than a 4-bit Qwen 3.5 27B. The consensus that emerged is that this is a tradeoff, not a clean win.
Here is the honest caveat, and Meta says a version of it itself. Every benchmark claim in the launch is first-party. Meta published an evaluation methodology document describing a consistent harness, averaged runs, and best-effort reproductions of competitor baselines -- more disciplined than a typical launch post, but still Meta grading Meta. No independent benchmark writeup confirming the agent claims has surfaced. And Meta's own model card says Glimmer is "generally less capable" than frontier-class models, with safety training aimed at refusal behavior and resistance to prompt injection. That framing is the tell: this is a guardrailed consumer-device agent, not a bid for the frontier. Read alongside Meta's earlier Muse Spark agentic stack, it reads as a deliberate split -- the big model stays hosted, the small one comes home. Whether that beats the alternatives on your own workload is still something only you can test, and the fit numbers get tight fast once you feed it images or a long context.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)