Meta is handing developers the settings needed to run and customize Muse Spark Open themselves. The company describes the model as geared toward coding and supporting AI agents, and notes it is less capable than Muse Wonder introduced earlier in the week. You can grab the weights from Amazon SageMaker AI, Google Cloud Model Garden and Hugging Face, and spin it up through partners like Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta also teamed with Apple so the model can operate on two of Apple's M5 Ultra Mac Studio desktops, giving developers a way to work on their own machines instead of relying only on remote servers.
Meta says releasing a less advanced open-weight model satisfies its earlier commitment to make its tech widely available while keeping the top-of-the-line models restricted for safety and capability reasons. As Chief AI Officer Alexandr Wang put it, "Because of both safety and capabilities, we will generally not open-source our most powerful models." He added that Meta intends to keep releasing strong open-weight options over time. The company is also building a cloud business where it would offer paid access to AI models and compute infrastructure.
Beyond open weights, the company is planning a cloud arm that sells access to its models and computing resources. According to Wang, releasing Muse Spark Open at no cost serves as the top of the funnel: developers can tinker locally first, and later purchase additional services from Meta once their projects require greater compute.
When technology and markets evolve quickly, stay grounded with our free Market Briefs newsletter that explains it simply.
Chinese players such as Alibaba Group Holding Ltd. and DeepSeek have seen notable success with open-source models that developers can tweak, and their rapid uptake in the US has dialed up the rivalry between Washington and Beijing. Wang argues US companies should keep publishing open weights to maintain an edge: "Most of the world's AI will run on open models," he said. "They should be American." He also expects Chinese-based developers to study and build on Muse Spark Open, adding, "It might even be good if Chinese models adopt some of our principals into their models."
Meta's posture here didn't come out of nowhere. Mark Zuckerberg had long championed opening up AI building blocks, but after Llama's open-weight models left Silicon Valley underwhelmed last year, he overhauled the AI team, hiring Wang, and in the aftermath that group delivered a run of stronger - but closed - models. Mark Zuckerberg long pushed to open up AI building blocks, but after Llama's open-weight models left Silicon Valley underwhelmed last year, he reshaped the AI organization and recruited Wang, and since then that team has turned out a series of more capable - yet closed - models. For everyday builders, that means a free on-ramp to experiment locally, with a path to more horsepower when projects outgrow a laptop.
Even amid complex AI and finance stories, sign up for free Market Briefs for a clear five minute daily digest.
