Meta launches on-device AI model to rival cloud-based systems

Technology giant Meta has unveiled Muse Glimmer, a new 30-billion-parameter open-weight AI model built by Meta Superintelligence Labs, designed to let users run persistent local agent workflows entirely on personal computers or Mac devices, as the company seeks to cement its position in the AI market and offer an alternative to cloud dependency.

Built to run seamlessly on local devices with or without an internet connection -- thereby ensuring data privacy and eliminating the latency of routing information to data centres -- Muse Glimmer requires more than 55 gigabytes of memory in full-precision mode, though Meta's 4-bit quantisation technology brings this down to under 20 gigabytes, making it viable on consumer PCs or Macs equipped with 24 or 32 gigabytes of GPU memory. Beyond standard chatbot functionality, the model is capable of autonomous agentic tasks, multi-step reasoning, multimodal input and analysis, local function calling, on-device coding, and LLM-as-a-judge evaluation, and has been trained on data spanning more than 100 languages. Its weights have been released under the Apache 2.0 licence on Hugging Face, with availability on Ollama, LM Studio and Unsloth to follow shortly; Meta is also working with AMD, Arm, Dell, Intel and Nvidia to optimise performance across devices, alongside forthcoming integrations with llama.cpp, MLX and ExecuTorch. According to Meta's own benchmarks, Muse Glimmer outperformed Google's Gemma 4-31B and Alibaba's Qwen 3.6-27B across all 12 comparison categories, with the company signalling plans to soon release weights for a more advanced model, Muse Spark 1.2.

Alongside the launch, Meta CEO Mark Zuckerberg issued a forceful call for changes to US AI policy and reduced restrictions on open AI models, arguing in a 14-page essay that artificial intelligence should be widely distributed rather than concentrated among a handful of firms. Citing Reuters, the article noted that Chinese firms -- including Moonshot's Kimi K3, Alibaba's Qwen 3.8-Max and DeepSeek's V4-Flash -- are mounting serious competition against leading US systems, even as flagship models from OpenAI, Anthropic and Google remain largely closed or tightly controlled. Zuckerberg argued that US companies must ease onerous data-use rules, regulatory complexity and restrictions on distillation techniques to maintain global leadership, adding that restricting access to foreign open AI models was not a viable solution.

Read More: Google advises job seekers to bypass its own AI hiring tool

Amid heightened industry concern following recent cybersecurity incidents involving various models and an attack on the Hugging Face coding platform, Meta announced that authority over approving safety standards prior to model release will be vested in the company's independent directors. The company is also continuing to pour massive investment into AI infrastructure, with spending this year expected to reach as much as $145 billion, while Zuckerberg separately announced a $1 billion fund aimed at ensuring local communities are not adversely affected by data centre construction.