Meta's 30B Muse Glimmer fits a local coding agent under 20GB, Apache 2.0

Meta Superintelligence Labs released Muse Glimmer today, a 30-billion-parameter model aimed at local agent workflows, with open weights on Hugging Face under an Apache 2.0 license. The pitch is narrow and concrete: a model small enough to run on a Mac or PC with a single consumer GPU, covering local agents and function calling, local coding, and LLM-as-a-judge evaluation. Optimized integrations for llama.cpp, MLX and ExecuTorch are described as landing in the coming days, not today.

The number that decides whether this is relevant to you is 20. At full precision a 30B model needs more than 55GB of memory. Meta quantizes the weights to approximately 4-bit precision, which brings the language model under 20GB and, by their account, leaves enough room for the KV cache, the perception encoder for image input and the speculative decoding drafter to sit alongside it in a 24GB or 32GB envelope.

That is the whole engineering argument compressed into one budget line. If it holds on your hardware, a full agent loop — planning, tool calls, retries, screenshot reading — runs without a network round trip and without a per-token bill.

The important detail for anyone sizing a deployment is that Meta is not quoting a weights-only figure and calling it done. The 24GB or 32GB envelope is explicitly meant to hold four things at once, which is the difference between a benchmark demo and a working agent.

Meta says it validated that the compression introduces minimal to no degradation on agentic tasks. The post does not publish the quantized-versus-full-precision numbers behind that statement, so anyone replacing a closed API with this should treat quality parity as a claim to verify on their own task suite rather than a settled result. The mid-training phase is described as using longer-context, agent-heavy data, but no context window length is given in the material available.

Meta frames latency as the thing that decides whether a local agent is usable at all: an agent that takes minutes to reply or plan its next step breaks the flow of real work. The fix is a lightweight drafter model based on DFlash that proposes blocks of tokens rather than emitting one at a time, which is where speculative decoding usually pays off — long reasoning chains and multi-step tool calls.

What the announcement does not contain is a throughput figure. There are no tokens per second, no acceptance-rate numbers for the drafter, no time-to-first-token, and no named GPU in the section describing these optimizations. "Fast enough to feel responsive" is a design goal in the post, not a reported measurement. Teams deciding between this and a hosted endpoint will have to benchmark the quantized model plus drafter on the specific card they intend to buy, because that number is the entire cost case.

The capability list reads like it was written by people who have watched small models fall out of agent loops. Muse Glimmer is trained and evaluated on end-to-end task completion, tool calling with precise schemas across extended workflows, multi-step reasoning over long horizons, and failure recovery — specifically, diagnosing an error and retrying when a tool call fails or returns something unexpected, rather than halting. Anyone who has run a 7B model against a real toolchain knows that last behaviour is the one that separates a usable agent from a transcript full of dead ends.

Controllable effort is the underrated line item for local serving. On a fixed 24GB card you cannot scale out to absorb a hard problem, so the only lever is spending more tokens on it. A model that exposes reasoning strength as a knob lets an orchestrator route cheap edits at low effort and reserve long chains for the cases that need them.

Meta states the weights are open sourced under a permissive Apache 2.0 license. That is the material difference against a closed API: the artefact is yours to run, and there is no per-token meter or rate limit between your agent loop and the model. The post also says Muse Glimmer was evaluated under Meta's Advanced AI Scaling Framework and assessed for open-weight release across all relevant categories. It does not describe any additional acceptable-use terms, usage thresholds or redistribution conditions layered on top, and the material here does not let me assert that none exist — read the Hugging Face repository before you assume. Coverage of the release framing appeared the same day in the Financial Times and Fox Business, whose headlines address Meta's return to open models rather than the deployment details.