Verdict, October 17, 2026. Mid for a hosted API or a coding-agent swap. OP only if you already have the cluster or the Mac Studio pair, the prompts cannot leave, and you can live with the license. OP or Mid has not run Muse Spark Open.
What shipped on October 15
On October 15, 2026, Meta released the weights for Muse Spark Open. VENDOR CLAIM The post calls it "the most capable US open-weight model to date." Source: Meta (opens in a new tab).
The model card, same day: 816B total parameters, 42B active, mixture-of-experts, 1M-token context. Input is text, image, video, and audio. Output is text. Checkpoints are BF16 and NVFP4. Source: model card (opens in a new tab). Transformers, checked October 17, puts that context at 1,048,576 tokens and routes each token to 8 of 256 experts. Source: Transformers (opens in a new tab).
Meta's deploy guide says NVFP4 serves at TP4 on 4x B300, and BF16 at TP8 on 8x B300. Source: Run inference (opens in a new tab). The post names the hosts: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Source: Meta (opens in a new tab).
Hosted API: Mid
PRICE Checked October 17, 2026, Together AI, Baseten, Fireworks, and Nebius Token Factory post $1.25 per million input tokens and $4.25 per million output tokens. Together, Baseten, and Fireworks also post cached input at $0.15. Sources: Together post (opens in a new tab), Together pricing (opens in a new tab), Baseten (opens in a new tab), Fireworks (opens in a new tab), Nebius (opens in a new tab), catalog (opens in a new tab).
Fireworks lists a 1040k context, text and image input, and a limited preview through October 29, 2026 on US-only capacity, with the usual 50% US-only premium waived. Nebius's October 16 post says text, image, and video on FP4. The catalog lists 1048k in us-north1 and no cached price. OpenRouter posts the same $1.25 / $4.25, cache read at $0.15 on Together, Fireworks, Baseten, and Modal, none on Nebius, a 262,144 completion cap, and no EU in-region route. Groq is absent there. Source: OpenRouter (opens in a new tab).
NERF Groq publishes no price. Access is a limited preview that needs organization approval. Context is 131,072 tokens, not 1M. Input is text and images, up to 5 images. Source: Groq (opens in a new tab). Meta's deploy guide states the same cap. Source: Run inference (opens in a new tab).
NERF Modal's library page, checked October 17, describes the model and prints no token price. Source: Modal (opens in a new tab).
Set that next to prices already on the Frontier Models board and the tools directory. Gemini 4 Argon is $2 / $10 per million through November 30, 2026. Haiku 5.5 is $0.10 / $0.50 up to 100K tokens. Muse Wonder is $4 / $20, and Wonder Full has no API.
VENDOR CLAIM Meta's model-card scores, all at reasoning effort xhigh, do not show a coding lead. Artificial Analysis Intelligence Index: 39, behind GLM 5.3 at 45, Kimi K3 at 44, and Qwen3.8-2.4T-A95B at 40. Terminal-Bench 2.1: 76.4, behind Kimi K3 at 85.0, GLM 5.3 at 83.9, Qwen3.8 at 82.0, and Beam at 80.1. The same table puts Spark Open first on MCP-Atlas (89.2) and JobBench (61.8). Those figures are Meta's. OP or Mid has not rerun them. Source: model card (opens in a new tab).
Default reasoning effort on the card is xhigh, the same setting as those scores. Low or medium is the shorter reply. Nebius warns that a small output cap can be eaten by reasoning before an answer appears.
Solo dev or small team on an API key: Mid. This is not a coding agent, so it does not move Codex, Claude Code, Antigravity, or Cursor on the AI Coding Agents list.
Your own machines: OP only if the box is already yours
NERF The model card's floor is at least 700GB of VRAM for NVFP4, or 2TB for BF16, on SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth, or llama.cpp. Source: model card (opens in a new tab).
NERF One Mac does not hold the NVFP4 checkpoint. The deploy guide says two M5 Ultra Mac Studios with 512GB each, or four with 256GB, plus macOS 26.2 or later, MLX 0.30.1 or later, Thunderbolt 5 between every pair, RDMA on, and a full NVFP4 copy on each disk. Four Macs are the fast setup. Source: Run inference (opens in a new tab). That is the blog's "two Apple M5 Ultra Mac Studios" line. Source: Meta (opens in a new tab).
NERF Smaller GGUFs stay large. The same guide quotes Unsloth at about 300 to 315GB for 2-bit and 670 to 700GB for 6-bit. BF16 language-model weights alone are about 1.63TB. Transformers NVFP4, checked October 17, installs from the finegrained-per-scheme-demo branch. Source: Transformers (opens in a new tab). A normal desktop is Mid.
Privacy and on-prem: OP when the cluster is the point
If prompts cannot go to a host, and you can feed 4x B300 for NVFP4 or the card's 700GB floor, the checkpoint is the feature. OP, and we have not run it. Internal use is carved out of the hosting clause below.
If on-prem means one GPU under that 700GB floor, it is Mid. The official weights do not fit, and a 2-bit GGUF still asks for about 300GB.
Building a product: OP inside your app, Mid if you are the API
The Muse Model License on the model card is dated October 15, 2026. Source: model card (opens in a new tab).
NERF You have to prominently display "Built with Muse Spark Open" in the UI of each product that uses the weights or a derivative, or in the docs, the website, or the settings. Derivatives include a model you fine-tune, distill, or improve using the weights or their outputs.
NERF If you or your affiliates serve model outputs to a third party through an API, and aggregate gross revenue exceeds $20 million in any consecutive 12 months, you must notify Meta and sign a separate agreement before offering these weights or derivatives as that service. The clause does not cover internal use, research, other business that is not model-as-a-service, or using the model inside your own product when that product is not a model API.
The usage policy on that page says the model is not intended for anyone under 18.
A startup that embeds Spark Open in its own app, and shows the notice: OP on the license, still Mid on quality until a test exists. A startup whose product is the API, once that $20 million line is crossed: Mid until the agreement exists. Muse Wonder is a different model, the October 12 API. Wonder Full has no API.
Where it sits
As of October 17, 2026, this article does not move a rank. S-tier on Frontier Models stays locked. Gemini 4 Argon, GPT-6, and the Claude 5.5 family stay A-tier on changelogs and prices. Muse Wonder stays B-tier.
Editorial call, not a board edit: Spark Open would sit beside Wonder in B. Meta's own table does not put it first on the intelligence index or on Terminal-Bench 2.1. The coding-agent list ranks harnesses, and a weight file does not take those slots.
Loadout
- Mid: a hosted API, a normal desktop, or a product that is itself a model API once revenue clears $20 million.
- OP: on-prem, and you already have 4x B300 or the two-Mac M5 Ultra pair. Show "Built with Muse Spark Open" if users can see the product.
- Do not read October 15 as an S unlock or a coding-agent patch.
The weekly patch is on beehiiv: subscribe free (opens in a new tab).