Meta has released a 30-billion-parameter model with a more useful target than gaining another couple of benchmark points: run a complete agent on personal hardware.
Muse Glimmer is released as open weights under Apache 2.0, according to Meta. It targets tool calling, coding, longer agentic tasks and multimodal text-and-image input.
The more interesting specification is its memory budget.
The quantized language model drops below 20 GB
Meta estimates that a full-precision 30B model would require more than 55 GB of memory for the weights alone.
Glimmer uses roughly 4-bit quantization and Meta says the resulting language model fits in under 20 GB. Its performance measurements include a 17 GB K-Quant build.
That is not the entire inference footprint. An agent still needs KV cache for working context, a visual perception encoder and, in Meta's proposed setup, a small companion model for speculative decoding.
The company targets a 24 or 32 GB envelope for the complete stack.
That is where a 30B model becomes genuinely interesting for local work. Seventeen gigabytes is not exactly tiny unless your definition of tiny owns a very expensive laptop, but it does fit hardware already sitting on the desks of some developers and creators.
Glimmer is trained to do more than answer
Meta describes Glimmer as a compact agent-oriented member of its Muse family.
It is trained for structured tool calls, multi-step planning, recovery after failed tool calls and work inside agent scaffolds. A separate perception encoder lets it consume interleaved text and images.
Meta publishes results on SWE-Bench, τ-Bench, MCP-Atlas and other agent-oriented evaluations.
Those are still Meta's own measurements on release day. They tell us what the company observed under its test conditions, not how the model behaves in a messy seven-year-old repository with strange tooling and decisions nobody remembers making.
The useful point is not to declare a new winner. It is to look at what kind of workload Meta is trying to make practical locally.
A smaller model helps it generate faster
Glimmer ships with a lightweight drafter based on DFlash.
With speculative decoding, that smaller model proposes blocks of future tokens. The main model verifies those proposals in parallel, accepts the correct ones and corrects the rest.
Meta says this speeds generation without changing the main model's output quality. Quantized drafter versions are provided to reduce their additional memory cost.
The company measured its 17 GB K-Quant configuration with the drafter on M4 Max and M5 Max MacBooks and an RTX 5090.
That still does not produce one universal tokens-per-second number. Context length, cache size, quantization and hardware all matter. The design goal is simpler: a local agent cannot feel interactive if every tool call gives you enough time to reconsider your career.
The weights are available; the easy integrations are still arriving
Meta says the weights are available on Hugging Face now.
The rest needs more careful wording. Optimized llama.cpp, MLX and ExecuTorch integrations are described as arriving in the coming days. The announcement also names Ollama, LM Studio and Unsloth as upcoming partner paths.
So “Glimmer already runs everywhere” would be premature.
The concrete release today is the weights and the hardware target. The easy experience in common local tools depends on integration work that follows.
Local AI is becoming an agent architecture problem
Local models have often been compared like smaller cloud chatbots: tokens per second, memory size, reasoning score.
Glimmer shifts the constraint slightly. Memory has to hold more than weights. It needs visual perception, working context and a decoding companion while the model calls tools and recovers from errors.
The PC is no longer being asked to host only an LLM. It is being asked to host a small agent stack.
The next useful test is less tidy than Meta's charts: what those 17 to 20 GB actually feel like when attached to real files, real tools and a user who has no interest in turning their computer into a permanent benchmark rig.