Meta's Muse Spark Open is out - open weights at $1.25/$4.25, with a license clause to read first

Meta released Muse Spark Open on 15 October: an open-weight model with 816B total parameters (42B active per token), a 1M-token context window, and a stated focus on coding and agent work. Meta calls it "the most capable US open-weight model to date." The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs are live at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Read that headline claim carefully. The qualifier "US" is doing real work.

What is actually new

Meta's best models, Muse Wonder and Muse Wonder Full, stay closed. Chief AI Officer Alexandr Wang said plainly that Meta "will generally not open-source our most powerful models." Muse Spark Open sits a tier below. What is new is that a US frontier lab has shipped weights in this size class at all, with broad hosting from day one.

On capability, Meta's own model card is more modest than the launch copy. Muse Spark Open posts the top score in its comparison table on MCP-Atlas, a tool-use benchmark, at 89.2. On Terminal-Bench 2.1, which tests agents working in a command line, it scores 76.4. Z.ai's GLM 5.3 scores 83.9 and Moonshot's Kimi K3 scores 85.0 in the same table. So the strongest open models for agent work are still Chinese. Muse Spark Open is the strongest one you can pick if a US origin matters to you, your customers or your compliance team.

All of these numbers come from Meta's evaluations. Nobody has independently tested it yet.

What it costs to call

On OpenRouter it is priced at $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million for cached input. That is the same rate OpenRouter lists for Meta's closed Muse Spark 1.3. For comparison, Meta's closed Muse Wonder lists at $4 / $20 on the same site, and DeepSeek's open V4.1 Flash 1017, released two days later, lists at $0.15 / $0.60 outside peak hours.

So the open weights do not buy you a cheaper API. They buy you something else.

What it means for a business owner

The real change is control. With a closed model, the vendor decides the price, the deprecation date and what happens to your prompts. With open weights, you can move the same model between seven hosts, or run it inside your own AWS, Google Cloud or Azure account so customer data never leaves it. If a host raises prices, you switch hosts. Your prompts, evaluations and tuning keep working, because the model behind them is identical.

That matters most for automation that touches regulated or sensitive data: contracts, HR files, patient or client records. It also matters for anything you plan to run for years. A support triage agent or an invoice-processing pipeline built on a closed model inherits that vendor's roadmap. One built on open weights you control does not.

The 1M-token context is useful for the same work: long contracts, full case files, a whole codebase in one call.

The honest caveats

**You will not run this in the office.** Meta says it is optimized to run on two Mac Studios with M5 Ultra chips. The model card asks for at least 700 GB of GPU memory for the compressed NVFP4 weights and 2 TB for full precision. For most businesses "self-hosting" means renting GPUs in a cloud account, and at low volume that costs more than the API. It pays off only at steady, high volume or when data rules leave no other choice.

**Read the license.** The Muse Model License allows commercial use, fine-tuning and resale. Two terms catch people. First, you must display "Built with Muse Spark Open" prominently in your product, documentation, website or settings. If you are putting an agent in front of your customers, that attribution goes with it. Second, if you resell API access to the model and your group's revenue passes $20 million in any 12 months, you must sign a separate agreement with Meta. Internal use and powering features in your own product are exempt from that clause.

**It is not the best model for every job.** If your workflow runs inside a terminal or a desktop, Meta's own table shows open alternatives scoring higher. If you have no data-residency constraint, a closed model may still be the better tool.

**Meta adds its own warning.** The model card says the open weights received extra safety training but should still run inside a system with its own guardrails. That is true of every agent. Log every action, cap what it can spend or send, and keep a human on anything irreversible.

What to do about it

Do not migrate anything yet. Pick one workflow where data control is the real blocker, the one your team keeps saying "we can't send that to OpenAI" about. Run 30 to 50 real cases through Muse Spark Open on a hosted API and through whatever model you use now. Compare the outputs and the bill. If the open model holds up, you have a path to running that workflow inside your own cloud account later, on terms no vendor can change under you.