Muse Spark Is Now an Open-Weights Model

Meta released the weights for Muse Spark Open, describing it as the most capable US open-weights model to date. The weights are available today on Hugging Face under the Muse Model License, and the model is tuned to run on the SGLang and vLLM inference engines with optimization work covering NVIDIA, AMD, and Intel hardware. The release builds on the August launch of Muse Spark, and this open-weights version adds safety post-training intended to make the model more robust against misuse when it is deployed standalone without system guardrails. Meta characterizes the result as a top open source choice for safety.
Muse Spark Open is an 816B total, 42B active parameter model with a context window of 1M tokens. That mixture-of-experts shape is worth unpacking: a large total parameter count gives the model broad stored capacity, while only a fraction of those parameters are active per token, which keeps inference compute closer to what a much smaller dense model would cost. The 1M-token window is the other headline number, and Meta frames it as enough to hold an entire project, including complex dependencies, legacy code, and thousands of files, without splitting a task into pieces. The model is natively multimodal and shares the same base model as Muse Spark 1.3. Meta positions it as coding-optimized, trained specifically for agentic coding workflows, with claimed strengths in first-attempt accuracy, reliable tool calling, and codebase understanding across repositories developers actually work in. Training reportedly drew on real developer workflows such as multi-file refactors, debugging sessions, and long-running tasks.
On the multimodal side, Muse Spark Open handles image and document understanding, and it also has audio and audio-video perception. Meta describes it as a reasoning model that processes image, audio, and text inputs natively. This is a common direction for frontier agentic models, where the ability to read a screenshot, hear audio, or parse a document in the same loop as code generation reduces the glue code a developer has to write.
The agentic design is the other major theme. Muse Spark Open maps out steps, acts, observes, and self-corrects over long horizons rather than making single-shot tool calls. It is trained to browse the web, including multimodal browsing, and to operate through an agentic harness that can spawn and coordinate subagents for decomposable work. It also has computer use capabilities unified across desktop, mobile, and browser, and it has been trained on long-horizon, multi-site tasks. One practical detail for teams building on top of it: the model generates separate commentary that lets it state intent before acting, so a caller can see what the model is doing and why. This suggests a design aimed at auditability inside agent loops, where knowing the plan before a tool call matters for both debugging and permissions.
Meta's stated motivation is open source support for the ecosystem. Because the company maintains PyTorch, it used Muse Spark Open as a showcase for two PyTorch projects. The first is Helion, described as a PyTorch-native high-level DSL for performant and portable ML kernels. Helion is a kernel authoring language aimed at writing high-performance kernels across heterogeneous hardware; it became part of the PyTorch Foundation in April 2026 after early adoption in the open-source community. Muse Spark Open ships with Helion implementations of operations such as RoPE, RMSNorm, and quantization, reported to deliver 1.03 to 3.95 times speedups versus custom kernels written in hardware-specific languages such as NVIDIA CUDA and CuTe DSL on NVIDIA GB300. Those same kernels were directly portable to AMD, scoring similar performance improvements on AMD MI350 with an overall 3 percent improvement in end-to-end throughput. Meta also launched Helion megakernels for Muse Spark Open in SGLang: these fuse multiple operations into a single persistent kernel to improve low-latency inference and can be autotuned for different shapes and hardware. Users can further autotune Helion kernels across accelerator architectures including NVIDIA, AMD, and Google TPU, along with quantization formats, tensor shapes, and parallelism strategies, without maintaining separate hand-tuned implementations for every configuration.
The second PyTorch integration is TorchTitan, a PyTorch-native platform for rapid experimentation and large-scale training of generative AI models. Starting today, TorchTitan adds a model definition for Muse Spark Open so developers can tune it to their needs. It includes native support for multi-turn supervised fine-tuning and LoRA out of the box, plus reinforcement learning capabilities through TitanRL for fine-tuning the model.
Meta also points to earlier model family releases: Muse Spark is the full-size model in the family, Muse Glimmer is a small dense model released previously, and ExecuTorch is available for getting started with it.
Why it matters: teams that want a large frontier-class model they can host, inspect, and fine-tune now have an open-weights option with a 1M-token context, native multimodal input, and agentic behaviors aimed at long-running coding tasks. The Helion and TorchTitan integrations are arguably the more durable story for developers: they promise portable kernel performance across NVIDIA, AMD, and eventually other accelerators, which lowers the cost of maintaining hardware-specific inference stacks. The trade-offs are the ones that come with any 816B total model: substantial memory and serving requirements, and a license that teams will need to review before commercial use.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Today, Meta is releasing the weights for Muse Spark Open, the most capable US open-weights model to date. It’s available today on Hugging Face under the Muse Model License. Muse... The post Muse Spark Is Now an Open-Weights Model appeared first on PyTorch .