میوز گلمر 30B

Meta's open-weight 30B model, optimised for local, always-on agent workflows. It is a dense transformer with a dedicated perception encoder, distilled from Muse Spark and purpose-built to run autonomous agents on consumer hardware. Meta aims it at end-to-end agentic task completion, reliable tool use, multi-step reasoning, failure recovery and use as an evaluation judge, with text and image input and selectable reasoning strength.

The argument Meta makes for this model is about where it runs. Most agent deployments assume a network and a datacentre; an agent that manages your calendar, drafts your messages and organises your files needs deep access to personal context, and Meta's answer is to put the model on the machine that already holds it. Everything about the design follows from that constraint - a compact architecture, a distillation recipe that pulls agentic reasoning out of a far larger teacher, and inference work aimed at latency rather than peak quality.

Meta breaks the recipe into three named stages. The first trains Glimmer on Muse Spark's own outputs rather than on raw text alone, so the small model inherits the large one's distribution. The second stretches the context and tilts the corpus towards agent transcripts that show their reasoning, kept diluted with ordinary text. The third layers three techniques over one another - supervised fine-tuning, on-policy distillation, then reinforcement learning - and applies them to general, reasoning, coding and agentic work alike. Meta says the result was assessed for open-weight release under its own Advanced AI Scaling Framework across every relevant risk category.

The capability list is written for agent builders rather than chat users, and one entry stands out: what happens after a tool call goes wrong. Meta says Glimmer was trained to read the failure, work out why it happened and try again, instead of halting on the error - which is the difference between an agent that finishes and one that needs a human. Alongside that sit precise function calling with real schemas across extended workflows, multi-step planning, a dedicated perception encoder that lets it read interleaved screenshots, charts and documents, selectable reasoning strength, and training data spanning more than a hundred languages.

Making it fit was its own engineering problem, and Meta shows the arithmetic. Thirty billion parameters at full precision needs over 55 GB; quantised to roughly four bits the language model drops under 20 GB, which leaves room for the working memory, the perception encoder and a speculative-decoding drafter inside a 24 GB or 32 GB budget - with, the lab says, minimal to no degradation on agentic tasks. A lightweight drafter proposes whole blocks of tokens that the main model verifies in parallel, which Meta reports as a large speed-up on a high-end consumer GPU and a meaningful one on laptop silicon.

Meta's own comparison table is unusually honest about the shape of the trade. Against the two open models near its size it leads on the agent-completion and tool-calling columns - the ones it was built for - while trailing on terminal-style coding, on computer use, and on general knowledge and hardest-exam questions. Read it as a model tuned for finishing tool-driven tasks locally, not as a small general-purpose frontier model.