Summary: Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, a 30B agentic model under Apache 2.0 for local agents, function calling, coding, and LLM-as-a-judge tasks. Its 4-bit build fits under 16 GB and runs on consumer GPUs, with reported RTX 5090 speeds of 75–233 tokens per second. The release ends Meta’s open-weight hiatus, though its dense design is slower than sparse peers.
On August 10, 2026, Meta Superintelligence Labs introduced Muse Glimmer — a 30-billion-parameter agentic model whose open weights went up on Hugging Face under an Apache 2.0 license. The positioning is unusually narrow: this model was built for always-on local agent workflows, and it is small enough to run on a Mac or a PC with one consumer GPU. Meta names the intended jobs directly — local agents and function calling, local coding, and LLM-as-a-judge evaluation.
The release also closes a long gap. Meta had not shipped open weights in more than a year before this. Set against Llama 4’s flop, a restructured AI group, and critics openly questioning Mark Zuckerberg’s commitment to open source AI, the launch carries more weight than “yet another small model” ever would.
30B Parameters on One Consumer GPU
Run a 30-billion-parameter model at full precision and the memory bill tops 55 GB — past anything a consumer GPU can offer. Quantization is Meta’s way around it. As The Register’s analysis has it, the 4-bit build lands at just under 16 GB, so it fits entirely on 20-to-24 GB cards such as the RTX 3090 or 4090 and the RX 7900 XT/XTX. Anyone on a 16 GB card has to settle for a less accurate 3-bit quant.
The speed figures hold up well for local hardware. Meta reports that an RTX 5090, with 1.8 TB/s of memory bandwidth, reaches 75 to 233 tokens per second, the top of that range leaning on speculative decoding driven by a small draft model. On an M5 Max MacBook Pro, the estimate is a still-usable 26 to 58 tokens per second. Software kept pace: optimized integrations for llama.cpp, MLX, and ExecuTorch were promised within days of the announcement, while Ollama and LM Studio already offer downloads.
Two points stood out from Simon Willison’s early hands-on. The license came first — Apache 2.0, he called it “a step up from the janky Llama licenses of old,” which matters to anyone building the model into a product. The second was multimodality: he had it describe a photograph in detail, then aimed his coding-agent plugin at a fresh checkout of Datasette and watched it explore the codebase through tool calls to answer a question about how authentication works. His practical conclusion: on a machine with 32 GB of RAM or more, a model this size still leaves plenty of room for whatever else you are running.
Distilling Agentic Ability from a Teacher
What the announcement does not dodge is why a local agent is so hard to pull off. One that manages your schedule or organizes your files has to deliver long-horizon execution, precise tool calling, multimodal understanding, long-context memory, and instruction following — all at the same time, and all within the memory and compute budget of consumer hardware. Meta’s answer is a novel distillation recipe that carries agentic reasoning over from a much larger teacher model. That teacher, according to The Register, is Muse Spark, Meta’s proprietary flagship.
Meta’s own numbers put Muse Glimmer up against Gemma4-31B and Qwen3.6-27B, the obvious peers in this size class. The evaluations Willison quoted range further, taking in full-task benchmarks — DeepSearchQA, MCP-Atlas, τ-Bench, and SWE-Bench — which ask whether a model can actually operate inside scaffolds, write and debug code, and carry multi-turn requests through to completion instead of merely answering trivia.
Why Meta Came Back to Open Weights
The timing is hardly subtle. The Register frames it bluntly: Chinese open-weights models — Kimi K3, Qwen 3.8-Max, DeepSeek V4 Flash — have owned the AI news cycle for months, while American companies struggle with the absence of credible domestic open alternatives. Meta built its reputation on the Llama family, then went quiet after Llama 4 disappointed. Muse Glimmer breaks that silence.
A second announcement hid inside the same day. Alexandr Wang, Meta’s superintelligence chief, posted on X that open weights for Muse Spark 1.2 are coming “soon.” That model is Meta’s most capable, the first to meaningfully compete with OpenAI, Anthropic, and Google. Taken together, the two messages read as follows: Glimmer is the advance party, and the flagship’s open version is the main course. Whether that release keeps pace with the Chinese frontier is the real test of Meta’s renewed open-source commitment.
What Changes for the Local AI Ecosystem
Three threads are worth tracking. First, an open 30B model with genuine agentic training raises the floor for local coding assistants and personal agents alike — agents that need neither network access nor API keys move from demo to daily driver. Second, Apache 2.0 clears away the licensing friction that made enterprises hesitate around the old Llama terms, and community fine-tuning groups such as Nous Research will almost certainly ship derivatives. Third, keep the limits in view: being dense, Glimmer spends all 30 billion parameters on every token, so it is inherently slower than sparse designs like Qwen 3.6-35B-A3B. The Register also cautions that the benchmark comparison may not age well, with Qwen’s next 27B model expected imminently.