Meet the models. Models power AI products. They are different from an assistant like Muse or a connection service like OpenRouter. Explore six downloadable releases by use, license and hardware.
ONE NAME, DIFFERENT THINGS Where does Muse fit? Meta uses “Muse” across an assistant and several models. Their access differs.
THE ASSISTANT Muse The personal agent. Meta said it launched powered by Muse Spark. Meet Muse · Meta source
HOSTED MODELS Spark 1.3 & Wonder Spark 1.3 is available via Meta Model API, Muse Code and OpenRouter. Wonder is another hosted model, planned for the Muse app. Spark 1.3 · Wonder
DOWNLOADABLE MODELS Glimmer & Spark Open Weights are released, but their licenses and computing needs differ. See them below
Six to explore. Choose by the job, the license and the hardware you have—not by a single “best” score.
01
Meta · A local agent engine Muse Glimmer Explore an agent model on a capable computer, with tool use and image understanding.
The 30B model still needs substantial memory; a quantized version fits a 24–32 GB envelope. Meta research 02
Google · A range of sizes Gemma 4 Find a size that fits your project, from edge devices to larger local or hosted setups.
Choose a specific variant before estimating device requirements or running cost. Google Open Source 03
Ai2 · See how it was made Olmo 3 Study, adapt and reproduce a model with unusually complete training materials.
A good research starting point; pick a model size and check its deployment requirements. Ai2 research 04
Qwen · A builder-friendly option Qwen3.6-27B Experiment with a downloadable multimodal model for coding and general tasks.
A 27B model calls for a capable machine or hosted inference. Qwen model card 05
Mistral · A model for production builders Mistral Small 4 Explore multimodal, reasoning and coding work in one released model.
Despite its name, self-hosting needs server-scale hardware; a hosted option may be simpler. Mistral release 06
Meta · Frontier-scale open weights Muse Spark Open Explore Meta’s large model for coding, multimodal and long-running agent work.
Different from hosted Spark 1.3. Meta’s local example uses two M5 Ultra Mac Studios; commercial terms are limited. Meta research What does “open” mean here? Downloadable weights let you run or adapt a model under its license. They do not automatically make its training data or full development process open, and running the model may still cost money. We describe each release precisely rather than calling every model “open source.”
Read the Open Source AI Definition Sources checked 16 October 2026. Model access, licenses and hardware requirements can change.