Meta Opened 816B Weights, and the API Still Costs $4.25 per Million Output Tokens

AI 요약

• Meta published the weights for its 816B-parameter Muse Spark Open.
• Reselling API access to it requires a separate Meta agreement past $20M in revenue.
• Per-token pricing, $4.25 per million output tokens, is unchanged from before the release.

The repository meta-models/Muse-Spark-Open-816B-A42B went live on Hugging Face on October 15. Anyone can now download an 816B-parameter model and run it on their own servers. Measured against April, when Meta shipped its first model, Muse Spark, with the weights closed and walked away from open source, that is a reversal in six months.

What went out is not Meta's best model, though. And even with the weights open, the price of calling it through an API is exactly what it was before. The reason is written into the license.

Distribution changed, capability did not

A new name usually implies a better model. Meta wrote down that this one is not. The PyTorch blog states that Spark Open "shares the same base model as Muse Spark 1.3." Muse Spark 1.3 is the model Meta has been selling through its API since September 2. The one difference is extra safety post-training done on the assumption that the weights would be public. In Meta's wording, it was trained further to be "more robust against misuse as a standalone model without system guardrails."

That post-training shows up in Meta's own safety numbers. Refusal rate on cybersecurity jailbreak attempts is 99.2% on Meta's internal benchmark and 94.8% across public benchmarks combined. Kimi K3 measured under the same conditions scored 75.5% and 70.2%. Single-turn harmful requests in chemistry and biology were refused more than 98% of the time, and the model scored 98.1% on MASK, an honesty evaluation.

The spec sheet is on the model card. It is a mixture-of-experts model: 816B parameters in total, of which only 42B activate for any one token, with a 1M-token context window. It takes text, images, video, and audio as input but emits text only, so it will not generate images or speech for you. Reasoning effort has four settings (low, medium, high, xhigh), and the default is xhigh.

The load-bearing word in "most capable US model" is "US"

Meta introduced Spark Open as "the most capable US open weights model to date." Remove "US" and the sentence stops being true. In the table Meta published on its own model card, Spark Open loses on the composite score to two Chinese open weights models.

The split is clean. Spark Open leads on tasks that involve calling tools repeatedly to get work done. On MCP-Atlas, which measures choosing the right tool through an MCP server, it scored 89.2 against 84.2 for GLM 5.3 and 82.3 for Kimi K3, and it takes first place on JobBench, a bundle of practical work tasks, at 61.8. On Terminal-Bench 2.1, which has the model actually fix code in a terminal, it falls to 76.4 against 85.0 for Kimi K3 and 83.9 for GLM 5.3.

All of those figures are Meta's. VentureBeat noted that as of publication it had not found an independent measurement of Spark Open. In the Hacker News thread (440 points, 267 comments), AustinDev reported running it locally and finding it both slower and lower-scoring than GLM 5.3 Flash.

Three days earlier matters too. Meta announced Muse Wonder and Muse Wonder Full on October 12 and released weights for neither. The Artificial Analysis figures VentureBeat reported put Wonder at 52 and Wonder Full at 56, above Spark Open's 39. What opened is not the frontier but the tier below it.

Why free weights did not move the API price

The license is not Apache 2.0. It is the Muse Model License, written by Meta. You can use, modify, fine-tune, distill, and resell the model, with two conditions attached.

The first is attribution. "Built with Muse Spark Open" has to appear prominently in one of your product UI, documentation, website, or settings menu.

The second condition is the one that touches money. A business that sells access to the model itself as a service and whose revenue, summed with affiliates over the preceding 12 months, passes $20M must notify Meta and sign a separate written agreement before putting Spark Open or any derivative on that service. Internal use, research, and use as a component of your own product rather than as a model service are all exempt. Most companies embedding the model in their own application are unaffected. Only inference providers reselling tokens are caught.

The result is visible on the price list. Spark Open on OpenRouter is listed at $1.25 per million input tokens and $4.25 per million output. That is the same rate Meta charged for hosted access to Muse Spark 1.1 through 1.3. Normally when weights go public, several providers list the same model and undercut each other. That has not happened yet with Spark Open. Hacker News user popinman322 pointed at the separate-agreement clause as the cause: the larger a provider's revenue, the more likely it needs its own contract with Meta, which blunts the incentive for open providers to undercut each other and leaves efficiency gains as provider margin instead of lower rates.

The same Meta released Muse Glimmer 30B in August under Apache 2.0. Attaching conditions to the big model while fully opening only the small one is the same structure as Mistral Large 4 publishing weights without Apache 2.0.

Can you use it today

Three paths, with different thresholds.

For readers outside the US, the thing worth checking first is the absence of a clause rather than the presence of one. Neither the license nor the model card contains any country restriction. That is not a given in this category: the Tencent Hunyuan community license explicitly excludes the EU, the UK, and South Korea from its grant, so a team in Dublin or London cannot use those weights the way a team in Austin can. Spark Open carries no such carve-out, which means the license question for a builder in Singapore, Sydney, or Berlin is the same one a US builder faces: the $20M model-service threshold and the attribution line.

The API routes are Together AI, OpenRouter, Fireworks, Groq, Baseten, Modal, and Nebius. Beyond Hugging Face, the weights are also listed on Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, so if your data has to stay in a specific cloud region you can deploy it into one yourself rather than routing through a hosted endpoint. Neither Meta nor the listed providers published a region-by-region availability map, so if you are bound by data-residency rules, treat the specific serving region as something to confirm with the provider rather than something the model card answers. Under Singapore's PDPA, or the EU AI Act for anything shipping into the EU, self-hosting the weights in a region you choose is the path that keeps that question in your hands.

The self-hosting threshold is high. The model card asks for 700GB of accelerator memory at NVFP4 quantization and roughly 2TB at BF16. Only 42B parameters firing per token does not make this behave like a 42B model, because the active experts change with every token, so the full weight set has to sit in memory.

Serving options are broad. SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, and Apple MLX are officially supported, across NVIDIA, AMD, Intel, and Apple hardware. Lower-memory quantized builds are available through llama.cpp, Ollama, and LM Studio. For fine-tuning, PyTorch's TorchTitan covers supervised training, LoRA, and reinforcement learning. Meta says the Helion kernels it contributed to SGLang gave a 1.03x to 3.95x kernel-level speedup on NVIDIA GB300 and a 3% end-to-end throughput gain on AMD MI350.

Meta's "runs on a desktop" line comes with conditions of its own. The configuration pairs two M5 Ultra Mac Studios for roughly 1TB of unified memory, but the 512GB model is due late October and Apple has not announced a price. The M5 Ultra you can buy today starts at $5,499 with 96GB. On Hacker News, devy estimated Meta's configuration at $28,000 to $30,000.

Getting the weights is not frictionless either. The Hugging Face gated form wants your full legal name, date of birth, and full organization name with no abbreviations, and it cannot be edited after submission. Filling it in inaccurately can block your access to other models, not just this one.

If you already run GLM 5.3 or Kimi K3, Meta's own table does not give you a reason to migrate wholesale. It does give you a reason to pull out one tool-heavy workload and compare. Spark Open leads on exactly two rows in that table, MCP-Atlas and JobBench, and both measure choosing and using tools. Swapping the model ID to meta/muse-spark-open on OpenRouter and replaying the same prompts answers that within half a day. If you are weighing self-hosting, late October, when Apple publishes the 512GB Mac Studio price, is when the arithmetic is worth redoing.