Meta is back in the open weights business, though on a smaller scale than the reaction suggests. Muse Glimmer landed Monday on Hugging Face at 30 billion parameters. The number is unremarkable. The licence is not.
Glimmer ships under plain Apache 2.0, carrying no monthly-user ceiling and none of the bespoke community conditions that turned Llama's terms into a three-year argument among enterprise lawyers. For anyone who spent that period reading Meta's licence more closely than its benchmarks, this is the consequential change. It also settles a question left hanging when the company moved its flagship line behind closed weights, and it arrives with unrestricted rights to deploy, modify and redistribute.
The Memory Maths
Thirty billion parameters at full precision would want north of 55GB, which puts the release out of reach of every consumer card before a single optimisation is applied. So the shipping product is really the quantised one. Meta's 4-bit builds drop the language weights below 20GB, leaving room for the KV cache, the vision encoder and a drafter model to sit inside a 24GB or 32GB envelope.
Reported accuracy loss is 0.2 percent on the 32GB configuration and roughly 1 percent on the 24GB variant. Those are Meta's own measurements rather than independent ones, and worth reading accordingly. In practice the ceiling is an RTX 4090 or 5090, or an Apple machine with 32GB or more of unified memory; an ordinary 16GB laptop is still shut out.
Latency is handled by DFlash, a small drafter proposing sixteen-token blocks that the main model verifies in a single pass. Meta clocks 233.4 tokens per second on an RTX 5090 with it against 74.9 without, and 50.2 on an M5 Max MacBook Pro against 26.6. The floor matters more than the ceiling for most readers. Without a dedicated card, output drops to around a dozen tokens per second, too slow to sit through while an agent grinds through tool calls.
What Is Actually In The Box
The model aims at a narrower target than a chatbot. Meta Superintelligence Labs trained it against the loop an autonomous agent runs: form a plan, invoke a tool, read what comes back, keep going, and diagnose a broken call instead of stopping at it. That last behaviour is what separates a demo from something usable, and it is the part most small models get wrong.
The build is a dense transformer of roughly 29.6 billion parameters across 52 layers, paired with a 1.8-billion-parameter vision encoder for screenshots, charts and documents. Context runs past 131,000 tokens, language coverage exceeds 100, and reasoning effort is selectable across four levels. Meta lists OpenClaw among the supported agent scaffolds.
None of this was trained from scratch. Glimmer is distilled from Muse Spark, the proprietary flagship, using logit distillation, then pushed through agent-heavy mid-training and a post-training mix of supervised fine-tuning, on-policy distillation and reinforcement learning. Distillation is why a model this size can behave like one several times larger. It is also why its ceiling is set by a model the public cannot inspect.
A Crowded Bracket
Google's Gemma 4 and Alibaba's Qwen3.6-27B already hold this weight class, and Meta's own comparison against them reads as a split decision. Glimmer takes the agentic tests, posting 75.5 on MCP Atlas, 74.6 on DeepSearch QA and 51.2 on SWE-Bench Pro where Gemma manages 36.9. Qwen keeps OSWorld-Verified, TerminalBench 2.1, most of the multimodal measures and a narrow lead on SWE-Bench Verified.
One cost sits beneath those tables. Glimmer spends all 30 billion weights generating each token, where sparse rivals activate three or four billion, so throughput lags what the accuracy scores imply. Buyers comparing benchmark rows alone will misjudge how the model feels in use.
Outlook
Distribution arrived faster than the model did. Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter are live or close to it, with llama.cpp, MLX and ExecuTorch integrations queued behind them. Mark Zuckerberg has since promised to open the weights for Muse Spark 1.2, the larger foundation model Glimmer was distilled down from. That is the release with the power to move anything, and until it lands the licence remains the most quotable thing about this one. Glimmer, polished as it is, re-establishes a foothold rather than reclaiming ground from the Chinese labs that have run the open weights category all year.
