Meta used Muse Spark in April to show it could build a closed frontier model. On August 10, it moved part of the same Muse line back toward open weights: Muse Glimmer 30B is now on Hugging Face and is aimed at agents that run on a Mac or a consumer-GPU PC.
The point is not the parameter count alone. Open models from China and the U.S. have already made long context and benchmark claims ordinary. Meta is asking a more practical question: can a downloadable, commercially permissive model do long tasks, read images, call tools and retry after failures without sending every step to a cloud API?
“Some argue that superintelligence itself or a small set of experts who control it should decide what is best for humanity. We disagree.”
30B is a memory decision
The Hugging Face model card lists a roughly 29.6B-parameter dense causal Transformer with a roughly 1.8B-parameter ViT-G/14 perception encoder, 131,072+ token context, a 202,048-token vocabulary, training across more than 100 languages, and a January 4, 2026 knowledge cutoff.
The key number is memory. Meta says a full-precision 30B model would need more than 55GB. Glimmer uses about 4-bit quantization to push the language model below 20GB, leaving room for KV cache, the image encoder and a speculative-decoding drafter in a 24GB or 32GB envelope.
This is not a cloud latency metric or a token-price claim. It is a local deployment threshold. The company also ships a DFlash-based drafter that proposes 16-token blocks for parallel verification. Meta reports decode speed gains of 3.1x on RTX 5090, 1.8x on M5 Max and 1.5x on M4 Max for its K-Quant-17GB setup.
The benchmark story is agentic
Meta's materials focus on end-to-end task completion, reliable tool use, multi-step reasoning, failure recovery, multimodal input, OpenClaw-style scaffold compatibility, controllable reasoning effort and multilingual use.
The model card reports 75.5 on MCP Atlas Public, 74.6 on DeepSearch QA, 51.2 on SWE-Bench Pro and 76.0 on SWE-Bench Verified. On the multimodal side, it reports 78.8 on Charxiv Reasoning and 75.4 on ScreenSpot Pro, compared with Gemma4-31B Thinking Mode and Qwen3.6-27B Thinking Mode.
Those scores are not the same as enterprise task success. A real agent has to cross browsers, repositories, permissions, logs and databases. The meaningful shift is that Meta is measuring local models around tool loops rather than only around chat answers.
Open weights return with guardrails
Glimmer uses Apache 2.0. Meta says it can be downloaded from Hugging Face, with local paths through Ollama, LM Studio and Unsloth, edge support through llama.cpp, ExecuTorch and MLX, serving through vLLM and SGLang, and quick starts through Together AI, Fireworks AI and OpenRouter.
The contrast with Muse Spark is deliberate. Spark remains the stronger closed model and has moved into paid API access. Glimmer gives developers a local agent model instead of Meta's most capable system.
The model card also sets boundaries. Meta says Glimmer should be deployed with additional system guardrails, especially for irreversible actions, data minimization, scaffold boundaries and indirect prompt-injection resistance. A local agent can see screenshots, touch files and call tools, so running on the device does not automatically make it safe.
The next checks are concrete: real speed on 24GB and 32GB machines, quality of Ollama/SGLang/OpenClaw integrations, and how often users need to take over long tasks. If those hold, Meta gets a new open ecosystem entry after Llama. If they do not, Glimmer becomes another 30B model that is downloaded, quantized, benchmarked and quickly replaced.
Sources: Meta AI Research, Hugging Face model card, Business Insider, FoneArena, CocoLoop; verification covers 29.6B parameters, 1.8B perception encoder, 131,072+ token context, Apache 2.0 license, sub-20GB quantized weights, 24GB/32GB local deployment scope, DFlash 3.1x/1.8x/1.5x decode-speed claims, and MCP Atlas/SWE-Bench benchmark scope.