Meta Platforms has released Muse Glimmer, a 30-billion-parameter open-weight AI model built to run on consumer-grade hardware and offered on Hugging Face under an Apache 2.0 license, positioning it as an always-on local agent for coding, tool use, multimodal tasks and autonomous retry loops. Meta said the model was trained by a team led by Scale AI founder Alexandr Wang and distilled from the proprietary Muse Spark family through a three-stage process spanning pre-training, longer-context mid-training and post-training with supervised fine-tuning, online policy distillation and reinforcement learning. Architecturally, Glimmer combines a 27.9-billion-parameter text decoder with a 1.9-billion-parameter ViT vision encoder and a GELU projection layer, using a 52-layer Transformer and a context window extended beyond 128,000 tokens through sliding-window and grouped-query attention. Meta said the model supports more than 100 languages. To make local execution practical, Meta said full-precision deployment would require more than 55GB of VRAM, so it used roughly 4-bit quantization to compress the language model below 20GB and paired it with a DFlash-based speculative decoding draft model. On an RTX 5090, decoding speed rose to 233 tokens per second from 74.9 tokens per second, while an M5 Max MacBook reached 50 tokens per second, up 1.8x. Under SGLang with NVFP4 and DFlash, Meta said single-user throughput reached 236 tokens per second and batch throughput 1,452 tokens per second. In benchmarks, Glimmer scored 75.5 on MCP Atlas, ahead of Gemma4-31B at 54.2 and Qwen3.6-27B at 62.5 in tool-calling, but Meta's release also showed Qwen3.6-27B leading on harder computer-operation and terminal-coding tests including OSWorld-Verified, TerminalBench 2.1 and SWE-Bench Verified. Mark Zuckerberg used a same-day essay, "The Future is for Everyone," to argue that superintelligence should be distributed rather than concentrated and to explicitly cite DeepSeek, Qwen and Kimi as leaders in open-source AI. He also said Meta will hand model safety review authority to an independent board rather than leave release decisions solely with the CEO, while still supporting chip export controls to slow rivals. The release comes as attorneys general from California, Colorado, Kentucky and New Jersey seek $1.4 trillion from Meta in litigation the company has called a headline-grabbing number without precedent. Meta also highlighted broad day-one ecosystem support across llama.cpp, Hugging Face Transformers, Ollama 0.32.7, LM Studio and SGLang, while Alexandr Wang said an open-weight version of Muse Spark 1.2 is coming soon.