Meta launches Muse Spark Open: 816B open parameters with a restrictive license

Meta has introduced Muse Spark Open, an 816 billion parameter language model (with 42 billion active) that, according to the company, is the most capable open-weights model developed so far in the United States. The announcement comes with a one-million-token context window, available through the APIs of Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, and with downloadable weights from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and the Microsoft Foundry model catalog.

The news has two readings. The first is technical: Meta claims that Muse Spark Open surpasses the leading US open models, such as Inkling from Thinking Machines and Nemotron 3 Ultra from Nvidia, in its own tests, often by margins of 20 points or more. The second is legal and operational: the license is not as open as the name suggests.

A license with fine print

The model is distributed under the "Muse Model License," which Meta describes as having "limited commercial terms" that allow many developers to use the model for free. In practice, any company with more than 20 million dollars in gross revenue over a consecutive twelve-month period must notify Meta and sign a written agreement before offering the software or derivatives as part of a "Model as a Service" business. In addition, any product or service that uses the model or a derivative must prominently display "Built with Muse Spark Open" in its user interface.

This scheme is reminiscent of Moonshot's Kimi K3: internal use and integration into proprietary products are free, but there are conditions. The usage policy prohibits military and critical infrastructure deployments, an exception that Meta already introduced with Llama for the US government in 2024. Unlike Anthropic's usage policy for Claude, no abuse limits beyond that prohibition are specified here.

Performance: strong in agents, weak in coding

Meta has designed the model for "coding and agentic work," with special emphasis on long-horizon tasks, complex software engineering, and multi-step reasoning. In the company's tests, Muse Spark Open leads Kimi K3, GLM 5.3, and Qwen3.8 in agent and tool-use benchmarks, such as JobBench and MCP-Atlas, and roughly ties in SciCode. However, in pure coding it falls behind: in Terminal-Bench 2.1, Kimi K3 scores 85.0 and GLM 5.3 83.9 versus Spark Open's 76.4; in DeepSWE 1.1, the difference is 68 to 60. It is the same gap that Meta acknowledged when it launched the closed version of Spark in April, and it does not seem to have improved much since then.

For an SMB, these numbers translate into a practical question: is it worth betting on this model to automate development or support tasks? If the bulk of the work is code generation, the Chinese models still hold the advantage. If the focus is on agents that execute multi-step workflows, Muse Spark Open is an option to consider.

Can it be run locally?

An 816 billion parameter model is not intended for a laptop. Meta indicates that it can be run on two Mac Studios with an M5 Ultra chip and 512 GB of unified memory, linked together and using MLX. The company did so to run Muse Code and publish updates to the Mastodon client Ice Cubes. Apple does not currently sell that 512 GB model to the general public, but in a few weeks it will be possible to buy two machines with those characteristics for the price of a compact car. For an SMB, this is not a realistic deployment option, but rather a demonstration of capability.

The practical route for most companies goes through the APIs of the aforementioned providers or through cloud catalogs. There it is viable to test the model without investing in hardware.

What it means for your IT team

The arrival of Muse Spark Open reinforces a trend we have already been observing: large open models are increasingly competitive, but licenses are becoming more complicated. For an SMB's IT team, the implications are several:

• Cost and dependency: using the model via API avoids investment in infrastructure, but introduces dependency on external providers. It is advisable to evaluate the latency, cost per token, and data retention policy of each provider.
• License and compliance: if your company exceeds 20 million dollars in revenue, you need to review the license before integrating the model into a product you offer to third parties. If you do not exceed it, internal use is free, but the obligation to display "Built with Muse Spark Open" may affect brand image in proprietary products.
• Benchmarks vs. reality: Meta's numbers are good in agents, but weak in coding. Before adopting it, test the model with your own use cases. Do not rely solely on the comparison tables.
• Alternatives: if your priority is coding, you may be interested in exploring models such as Nvidia's, which has recently launched a 74B agentic model for laptops. You can read our analysis in Nvidia fits a 74B agentic model into 24GB for your laptop.

ForgeNEX recommendations

At ForgeNEX we recommend a prudent approach. First, evaluate whether you really need an 816B model for your tasks. In many cases, a smaller, specialized model may be sufficient and cheaper. Second, review the license with your legal team if you plan to integrate it into a commercial product. Third, test the model in a controlled environment with your own data and use cases before committing. And fourth, maintain a multi-provider strategy so as not to be tied to a single API.

The news also coincides with a moment of growing interest in security in the use of AI. If you are deploying agents or models in production, you may be interested in reviewing our recommendations on new AI decision models arriving in SOCs, where we address the risks of automating decisions without supervision.

In short: Muse Spark Open is one more step in the race of open models, but it is not a silver bullet. SMBs should look beyond the benchmarks and evaluate license, total cost, and real fit in their workflows.

Source: The New Stack. Analysis and adaptation: ForgeNEX.