Meta Returns to Open Weights With Muse Spark 1.2 & Muse Glimmer

Meta is releasing the weights for Muse Glimmer, a 30-billion-parameter dense model that can run on a laptop or a single consumer GPU, while preparing to open source a version of Muse Spark 1.2, one of its leading foundation models, in the coming weeks.

Muse Glimmer is being released under the Apache 2.0 licence, with its weights available through Hugging Face. Meta says the model is intended for local agents, function calling, coding and LLM-as-a-judge workloads without requiring a cloud connection.

For Muse Glimmer, integrations with llama.cpp, MLX and ExecuTorch are expected in the coming days. Meta also plans support through Ollama, LM Studio and Unsloth, alongside deployment options including vLLM and SGLang.

Meta said it is working with AMD, Arm, Dell, Intel and NVIDIA to optimise the model across different hardware platforms.

“Running models locally enables you to use AI anywhere, anytime, with or without an internet connection,” Meta said in its announcement.

Muse Glimmer can handle multi-step agent workflows, including tool use, failure recovery, multimodal inputs and long-context tasks. It supports text and image inputs through a dedicated perception encoder and can work with agent orchestration frameworks such as OpenClaw.

Meta said Muse Glimmer was trained using outputs from its larger Muse Spark model. The training process used logit distillation during pre-training, followed by mid-training on longer-context and agent-focused data.

For post-training, Meta combined supervised fine-tuning with on-policy distillation and reinforcement learning across reasoning, coding and agentic tasks.

The company said the approach was intended to transfer capabilities from a larger model into a model that can run within the memory limits of consumer hardware.

“At full precision, a 30-billion parameter model would require over 55 GB of memory,” Meta said.

Meta uses quantisation to bring the model down to under 20 GB at roughly 4-bit precision. This allows the model, its KV cache, perception encoder and speculative decoding system to operate within a 24 GB or 32 GB memory envelope.

Muse Glimmer supports tool calling, multi-step reasoning and failure recovery. Meta said the model can diagnose failed tool calls and retry rather than stopping the workflow.

It was evaluated on benchmarks including DeepSearchQA, MCP-Atlas, τ-Bench and SWE-Bench. Meta compared its performance with Gemma 4 31B and Qwen3.6 27B, saying Muse Glimmer performs strongly across several benchmarks for models in its size category.

The model also supports different reasoning strengths, allowing developers to trade off response quality and speed. It was trained on data covering more than 100 languages.

Meta is also using speculative decoding to reduce generation latency. Muse Glimmer comes with a lightweight drafter based on DFlash, which proposes blocks of tokens for the main model to verify in parallel.

“An agent that takes minutes to reply or plan its next step breaks the flow of real work,” Meta said.

The company tested its K-Quant-17GB version with a quantised DFlash drafter on Apple’s M4 Max and M5 Max Macs, as well as NVIDIA’s RTX 5090. Meta said the combination is intended to support real-time interaction while keeping processing on the device.