Meta Releases Muse Spark Open, an 816B Open-Weight Model

Meta has released the weights of Muse Spark Open, an 816-billion-parameter multimodal model that it calls the most capable US open-weight model to date. The release matters for two reasons: the model is large enough to compete with closed frontier systems on coding and agentic work, according to Meta, and it is built so that a well-equipped desktop, not a data centre, can run it. Independent verification of Meta's capability claims is still pending, and the license is not a conventional open-source one.

What Meta released

According to Meta's announcement, dated 15 October 2026, Muse Spark Open has 816 billion total parameters with 42 billion active per token, a 1-million-token context window, and native multimodality. Meta says it is designed for coding and agentic tasks, including long-horizon software engineering and multi-step reasoning. The active-parameter figure is the practical one to note: only a fraction of the model is used for each token, which is what makes serving it cheaper than the headline size suggests.

The weights are distributed under what Meta calls the Muse Model License, which it describes as having "limited commercial terms" while letting many developers use the model for free. That wording sits apart from the permissive licenses, such as Apache 2.0, that many developers associate with open source. Teams considering commercial deployment will want to read the actual license text, which this article did not review, before building on it. The term "open-weight" here means the parameters can be downloaded and run, not that the training data or full terms are unrestricted.

Distribution is broad. Meta lists the weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs from providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It also names support in inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA hardware. Fine-tuning is supported through PyTorch and TorchTitan.

The local-hardware angle

The headline deployment claim is that the model runs on a desktop. Meta says this works with two Apple M5 Ultra Mac Studios using the latest MLX framework, noting that each machine's integrated GPU can address up to 512 GB of unified memory and that the chip offers 1.2 TB/s of memory bandwidth. TechCrunch, in its report on the launch, cites Engadget for a requirement of at least 1 TB of RAM to run the model locally, and says Meta's best-performing setup is a Mac Studio with the M5 Ultra. Two 512 GB machines add up to that figure, so the accounts are consistent.

That is a very different bar from consumer hardware. A headline like "runs on Macs" should be read as "runs on a pair of top-configuration workstations", a cost level aimed at professional teams, labs and enterprises that need data to stay on-premises. Even so, it narrows the gap between frontier-scale models and local deployment, and it helps explain the emphasis on Apple's MLX in the inference-engine list.

What Meta claims, and what is not yet verifiable

Meta says the model outperforms other popular open-weight models, according to evaluation results it published. TechCrunch notes that its article gives no specific benchmark names or scores, and the portion of Meta's page reviewed for this article similarly pointed to a separate evaluation methodology report rather than listing numbers. Meta says a full evaluation report is "available soon" and that detailed results appear in the Hugging Face model card. Until independent evaluators reproduce the results, "most capable" should be treated as a company claim.

The only quantitative performance figures in Meta's post concern serving speed. It reports kernel-level speedups of 1.03x to 3.95x from custom Helion GPU kernels in SGLang on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD's MI350. Those numbers describe an inference optimisation, not model quality, and the modest end-to-end gain shows how much of the kernel-level improvement is diluted in a full serving stack.

Meta also showed demonstrations. According to TechCrunch, the model was paired with Muse Code, Meta's AI coding terminal, to fix bugs, add features and perform a full build on a Mastodon client, and it was used with DaVinci Resolve to review and crop footage. Meta's page describes on-device agentic workflows such as iOS code changes and video editing. These are vendor demonstrations rather than benchmarks, and they do not show how often the model succeeds across many tasks.

Safety and governance

Meta says the open-weight version includes additional safety post-training and that it met the release thresholds of its Meta Superintelligence Scaling Framework, which covers the most severe risks. Detailed safety results are said to be on the Hugging Face model card. Open release changes the safety picture because anyone can fine-tune the weights, including to remove safeguards, so post-training protections are a mitigation, not a guarantee. The usual debate over open weights, which weighs transparency and cost advantages against misuse risk, applies here as it does to other large releases. TechCrunch notes the common argument for open-weight models: they can be more cost-efficient and offer more transparency and flexibility than closed models.

Why it matters for democratised AI

Three things stand out. First, the distribution list suggests Meta wants the model available wherever developers already work, from the major clouds to specialised inference hosts, which lowers the switching cost for teams that already use those platforms. Second, the license choice keeps Meta in control of commercial terms even as the weights are public, which is a recurring pattern among large labs releasing open-weight models. Third, the hardware story pushes the idea that frontier-scale capability can run privately, which matters for organisations with data-residency or confidentiality requirements.

The open questions are practical. How well does the model hold up on independent coding and agentic evaluations? What do the license terms permit for companies of different sizes? And how well do community inference stacks, such as llama.cpp and vLLM, perform on a model of this size outside Meta's own setups? Meta's promised evaluation report and early third-party testing will be the first signals to watch, followed by whether developers actually adopt the model in production over alternatives.