On October 15, Meta released the weights for Muse Spark Open. The model is hosted on Hugging Face under the Muse Model License. It is the first open-weight release in the Muse Spark family: 816 billion total parameters, 42 billion active per token, a 1-million-token context window, and native support for image, audio and audio-video input. In its PyTorch blog post, Meta calls it “the most capable US open-weight model to date.”
The open version shares a base model with Muse Spark 1.3, which shipped the same day; the difference lies in post-training. The open release received an extra round of safety post-training, which Meta says makes it harder to misuse when deployed on its own, without system-level safeguards. Meta says that, under its own Superintelligence Scaling Framework, the model met the bar for release. The full evaluation report has not been published, and the company says only that it is “coming soon.”
Built for coding and agents
Meta's positioning is narrow: coding and agentic tasks. The training data includes multi-file refactors, debugging sessions and long-running tasks, and the company stresses first-attempt accuracy and reliable tool calling. On the agent side, the model was trained for long-horizon tasks that span multiple websites and can spin up sub-agents to divide the work, with computer use on desktop, phone and browser trained as a single unified skill. It also writes out a separate explanation of what it intends to do before it acts.
Meta's research blog has a Key Benchmarks section, but the text lists no scores, so third-party results will have to wait. Arena has already added the model to Agent Arena, with ratings not yet published.
Can two Macs run it?
Meta's desktop setup is two M5 Ultra Mac Studios, each with up to 512GB of unified memory, running the latest MLX. Apple's machine learning research team published its own measurements the same day. After NVFP4 quantization the weights come to about 493GB, and once the intermediate memory used during inference is added, a single machine cannot hold them. Two machines using pipeline parallelism reach 15.74 tokens per second; switching to tensor parallelism raises that to 24.99, and four machines with tensor parallelism reach 36.73, all at an 8,192-token test context. The machines are linked through the Thunderbolt 5 RDMA driver introduced in macOS 26.2, with two cables between each pair.
In the cloud it runs faster. SGLang and vLLM support it from day one, and Together AI, Modal and OpenRouter have all listed it. Together's serverless pricing is $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15, and it is free for a limited time until noon Pacific Time on October 18.
Meta is also using the launch to promote its PyTorch toolchain. The model ships with kernels written in Helion, covering operations such as RoPE, RMSNorm and quantization, which Meta says run 1.03 to 3.95 times faster on NVIDIA GB300 than hand-written kernels in CUDA and the CuTe DSL; the same kernels ported directly to AMD MI350 improve end-to-end throughput by about 3%. Fine-tuning is supported through TorchTitan, including multi-turn SFT, LoRA and reinforcement learning.
From closed to open
Put Meta's recent releases in a line and a pattern emerges. Muse Spark launched as Meta's first closed model, a break from the Llama-era approach. Version 1.2 in August arrived alongside the coding product Muse Code, and Meta had earlier released Muse Glimmer, a small dense model for on-device use. Now Spark itself has open weights, so Meta's model line runs a closed, iterating version and an open version side by side: the same foundation, separated by a layer of safety post-training.
The arrangement lets Meta cover both sides. The closed version keeps earning revenue in Muse products and the API, while the open version competes for developers and inference providers by securing day-one support, from SGLang and vLLM to several cloud providers. That part is already done.
The license remains an open question. Meta says only “limited commercial terms” and that many developers can use the model for free. The blog does not state the commercial threshold, nor whether it keeps the Llama license's cap based on monthly active users. The full terms are on the Hugging Face page, and any company planning to use the model commercially should read that license first.
Sources: PyTorch official blog, Meta research blog, CocoLoop, Apple Machine Learning Research blog, Together AI; Together's pricing page was used to verify per-million-token rates and the free period, and Apple's tests to verify the quantized size and tokens per second.