A 30-billion-parameter model sized for a single consumer GPU
Meta released Muse Glimmer on August 10, 2026, as an open-weight model that runs on a Mac or PC equipped with a single consumer graphics card. The model is available as a free download on Hugging Face under an Apache 2.0 license, which permits commercial use, modification, and redistribution without the usage restrictions attached to competing open-weight releases. Meta said the model targets agentic tasks such as schedule management, file organization, local coding, and function calling.
How the model fits on consumer hardware
At full floating-point precision, a 30-billion-parameter model would require more than 55 gigabytes of memory, exceeding any consumer graphics card on sale. Meta quantized the weights to approximately 4-bit precision and compressed the language model to under 20 GB, leaving headroom inside a 24 GB or 32 GB memory envelope for a KV cache, an image-processing perception encoder, and a speculative decoding drafter. Meta measured decoding speeds 3.1 times faster on an NVIDIA RTX 5090, 1.8 times faster on an Apple M5 Max, and 1.5 times faster on an M4 Max, with output quality matching standard token-by-token generation.
Training pipeline and benchmark results
Muse Glimmer is a distilled version of Meta's larger Muse Spark 1.2 model, launched as a closed flagship on August 5. Training ran in three phases: logit distillation from Spark, mid-training on longer-context and agent-heavy data, and a final stage combining supervised fine-tuning with on-policy distillation and reinforcement learning. Meta reported strong results for its size class against Gemma4-31B and Qwen3.6-27B on benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, which measure retrieval, multi-step agentic planning, and code writing and debugging.
Capabilities, tooling, and ecosystem support
The model is designed to diagnose failed tool calls and retry them rather than halt, accepts interleaved text and images including screenshots and documents, works with agent orchestrators such as OpenClaw, supports adjustable reasoning effort, and covers more than 100 languages. Meta said local operation allows on-device use without an internet connection and keeps queries from leaving the machine. Optimized integrations for llama.cpp, MLX, and ExecuTorch were due within days, while Ollama, LM Studio, Unsloth, Together AI, Fireworks AI, and OpenRouter supported the model at launch. Meta pointed larger-scale inference workloads to vLLM and SGLang, with hardware optimization work underway with AMD, Arm, Dell, Intel, and NVIDIA.
Open-source policy push and competitive framing
Alongside the release, CEO Mark Zuckerberg published a 14-page essay arguing that the U.S. should lower barriers to open-source AI development rather than centralize superintelligence. He said American open-source developers face regulatory disadvantages relative to Chinese competitors on issues including training data use and distillation techniques. Reuters, cited in coverage, reported that Chinese startups including Moonshot, Alibaba Group Holding, and DeepSeek have taken the front position in open-weight AI development with models matching leading U.S. systems, while the top offerings from OpenAI, Anthropic, and Alphabet's Google are not publicly available as open-weight systems. Reuters also reported Meta intends to release the weights of its flagship Muse Spark 1.2 as well, though that plan was not confirmed in Meta's own materials. Meta stock was up nearly 3% in premarket trading on the release day.
What remains uncertain
The exact memory footprint after quantization is described in the Quartz report as approximately 17 GB and in the BetaNews report as under 20 GB, a difference attributable to the two outlets' summaries rather than a stated Meta contradiction. Reuters' reporting that Meta intends to publicly release the weights of Muse Spark 1.2 is not corroborated by Meta's own materials in the source set, and the timing of that release, if confirmed, is not specified.
Share this article
