Muse Spark Open Is the Most Capable US Open-Weight Model. Six Chinese Models Score Higher, and the Mac It Runs On Is Not for Sale Yet.

Three days after Muse Wonder put Meta in the frontier tier with a model you cannot buy by the token, Meta released a model you can download. Muse Spark Open shipped today, October 15, 2026: "Today, we're releasing the weights for Muse Spark Open, the most capable US open-weight model to date."

Every word of that sentence is true, and the sentence is doing a lot of work. On Artificial Analysis's open-weights leaderboard, read today, Muse Spark Open scores 39 on Intelligence Index v4.3.2. Xiaomi's MiMo-V2.6-Pro scores 46, Z AI's GLM-5.3 45, StepFun's Step 5 Preview 44, Moonshot's Kimi K3 44, GLM-5.3-Flash 42, and Alibaba's Qwen3.8 2.4T A95B 40, which AA scores but leaves off that page. Muse Spark Open tied DeepSeek V4.1 Flash at 39 on the day this post shipped; DeepSeek's later Flash-1017 build was rescored to 45, see the October 17 and 18 update below. "US" is the qualifier that makes it first. AA's own article today, about NVIDIA's new Nemotron models, still calls Nemotron 3.5 Super VL at 26 "the leading US open weight model" and does not mention Meta.

The second claim in Meta's post is the one that will sell Mac Studios: Muse Spark Open "is the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios." We read the model card, the Muse Model License, Meta's evaluation methodology PDF, Apple's MLX distributed-inference post, Apple's Mac Studio page, and AA's boards. Here is what held up.

What You Are Downloading

The license is open-weight with strings, and AA's comparison pages get it wrong. The model card requires licensees to "prominently display 'Built with Muse Spark Open'" in the product, and anyone offering the model as a service above $20 million in revenue needs a separate agreement with Meta. AA's own model page says "commercial use allowed with restrictions"; AA's side-by-side comparison pages flag "Commercial Use Without Restrictions: Yes." Read the license, not the flag. Meta's previous open release, Muse Glimmer 30B in August, was Apache 2.0 and ungated; this one is neither.

Meta will not serve it to you. The Model API docs are explicit: "You don't call it over Model API." You download it or you rent it from a third party, and the third-party price, $1.25 and $4.25, is exactly what Meta charges for its closed Muse Spark 1.3, which scores 48 on the same index. Meta's own open model costs the same as Meta's better closed model.

Terminal-Bench appears twice at different versions. Meta's methodology PDF lists Terminal-Bench 4.0 as one of the ten AA index components on page two, then the card table reports Terminal-Bench 2.1 at 76.4%. AA's run of Terminal-Bench 4.0 scores Muse Spark Open at 8%. Those are different tests; never put them in one column.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the source and index version; Meta's numbers labeled as Meta's. All Meta figures are at xhigh.

Artificial Analysis, Intelligence Index v4.3.2, October 15. Muse Spark Open (xhigh) 39, tenth of 119 in its class, which AA calls "well above average among comparable models" with a median of 19, and "particularly expensive" among open-weight models of similar size. Blended price $0.78 per million tokens against $0.18 for DeepSeek V4.1 Flash at the same score. Output speed 102.7 tokens per second on the best host.

The pattern is consistent across the boards: Muse Spark Open is a knowledge-and-reasoning model that is weak at agentic and terminal work. On the agentic rows (Briefcase, GDPval, AutomationBench, Terminal-Bench) it trails both comparators; on exams and science it holds its own.

Now the vendor table. These are Meta's self-reported numbers from the model card, run at "the highest available reasoning effort for each model," with two caveats Meta prints in the methodology: Beam and Mistral Large 4 figures are self-reported by those vendors because their weights were unavailable, and the AA figures were pulled on October 7 with older version labels.

Meta's table shows Muse Spark Open leading on JobBench, MCP-Atlas, HLE-Diamond, AA-Omniscience, and three multimodal evaluations, and trailing Kimi K3 and GLM 5.3 on all three coding benchmarks, on GDPval-AA, on Toolathlon, and on GPQA. The JobBench lead over GLM 5.3 is 0.4 points.

Three honest observations:

• "Well ahead of Qwen 3.8, GLM 5.3, and Kimi K3" is Meta's table in The Decoder's words. The Decoder prefaces it with "Meta says" and its own verdict is "roughly on par with the most capable open-weight Chinese AI models." VentureBeat's is "a contender with specific strengths, rather than an overall open-weight leader." Nathan Lambert wrote that he had waited a long time for US open-weight models to be relevant again, and that there is still more work to do. Those are the independent readings.
• OpenRouter's listing calls it "a coding-optimized open-weight model. Trained specifically for agentic coding workflows." Meta's own table has it trailing on every coding row, and AA's Terminal-Bench 4.0 run gives it 8%. The listing text is a host's marketing, not Meta's claim.
• The safety numbers are the vendor's, and they cut both ways. On Meta's adaptive StrongREJECT v2, attack success is 0.5% against 62.9% to 86.4% for the compared open models. On FORTRESS, it over-refuses 17.4% of benign prompts, the highest in the table. No independent safety measurement exists yet.

The Two Mac Studios

Meta's claim: "Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512 GB per Mac Studio, so two machines can hold an entire model." Apple's own research post ran the numbers in NVFP4, and Apple's Mac Studio page carries the detail Meta's post does not: "512GB memory option for M5 Ultra coming late October."

So on launch day the two-machine desktop could not be ordered. What could be ordered was four 256 GB Mac Studios at $10,799 each, which VentureBeat priced at $43,196 and Engadget at roughly $38,000. Apple's measurements, at 8,192 tokens of context: 15.74 tokens per second on two machines and 36.73 on four, with time to first token of 8.4 and 4.8 seconds. At 65,536 tokens of context, first token takes 47.6 seconds on two machines in pipeline mode, 70.3 seconds in tensor mode, and 40.6 seconds on four. Decode at that context is 14.37 to 34.66 tokens per second.

That is a working desktop frontier-class model, and it is a slow one at long context. Engadget's headline put the other constraint plainly: you will need a terabyte of RAM to run it at native precision. For comparison, the hosted version is $1.25 and $4.25 per million with a free Together promotion through October 18.

Meta's developer post adds a speculative-decoding path, DFlash V2, at "3.7x throughput for a single request on AMD MI355X." Its code samples set reasoning_effort='low'.

Against Meta's Own Lineup

The comparison Meta's card does not make is the one a buyer makes first: how the open model stacks up against the closed ones Meta sells.

So the lineup is closed at the top and open below, with a nine-point gap between the open model and the closed model that shares its name and price. AA measures Spark 1.3 at 127.7 tokens per second on Meta's API against 102.7 for the open model on its fastest host. The open model's one structural advantage is the one that cannot show up in a table: it runs where you put it, under terms you can read, and nobody switches the version under you.

The lineage matters for one more reason. Several write-ups measured the jump from "Meta's last open model" as 30 index points, counting from Llama 4 Maverick at 9. Meta's last open release was Muse Glimmer in August at 17, so the gain is 22. Both are real gains; only one is the right baseline.

Where You Can Run It

Weights on Hugging Face (gated, with a form) and Amazon SageMaker JumpStart; Meta also names Google Cloud Model Garden and Microsoft Foundry, which we could not find on either vendor's docs today. APIs at Baseten, Fireworks, Groq (limited preview, contact sales, 131K context), Modal, Nebius, OpenRouter, and Together. Engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, MLX, with community GGUF, MLX, INT4, and MXFP4 builds already posted, several of them dated before the public launch. Not in Meta's consumer Muse app. Not on Meta's Model API.

Not yet on Arena.ai or Vals AI; it entered Arena's Agent and Code arenas today with no scores. Those print as dashes.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Muse Spark Open if you need a frontier-adjacent model with a million-token multimodal context that you can run on your own hardware under a license you have actually read, and your work is knowledge, exam-style reasoning, or MCP tool use, where Meta's and AA's numbers both put it ahead of its weight class. If you already own four 256 GB Mac Studios, this is the best US model that fits on them.

Skip it for coding and terminal agents, where it trails Kimi K3 and GLM 5.3 on every row and scores 8% on AA's Terminal-Bench 4.0; for anything that needs a fast first token at long context; and for hosted use, where DeepSeek V4.1 Flash scores the same for about a quarter of the blended price and Meta's own closed Muse Spark 1.3 scores nine points higher for the same price.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58; among open weights MiMo-V2.6-Pro at 46 and GLM-5.3 at 45 lead; Muse Spark Open at 39 is where the best US open model sits today, 22 points above Meta's own August release. Our model comparison has the current standings side by side.

Update, October 18, 2026

• October 17 and 18: DeepSeek replaced the model this post's table calls "DeepSeek V4.1 Flash (max) 39" under the same name. On October 17 deepseek-flash moved to a new build, DeepSeek-V4.1-Flash-1017, and on October 18 Artificial Analysis scored that build 45 on the same v4.3.2 index, six points above Muse Spark Open, at $0.26 per index task against $0.49 for the September build at the same list price. The 39, the 27% on Terminal-Bench 4.0, and the "same score for about a quarter of the blended price" comparison above all describe the September build, which now exists only as a separate Hugging Face checkpoint. Details in our DeepSeek V4.1 Flash 1017 post.
• October 16: The Hugging Face launch thread on Hacker News passed 848 points, with the debate centered on whether the $20 million clause and the attribution requirement make the model "open-weight," and on the Mac Studio bill. Together's free promotion runs until October 18 and Fireworks's limited preview through October 29. The 512 GB Mac Studio is still not orderable. Reflection AI's Beam, the only other US open model in Meta's table at launch, shipped Apache 2.0 weights the same day, after this post went up; Artificial Analysis has since scored Beam around 30, details in our Reflection Beam post. Artificial Analysis is still on index v4.3.2, with no Arena or Vals entry for Muse Spark Open.

Sources

• Meta: Introducing Muse Spark Open
• Hugging Face: Muse Spark Open model card and license
• Meta: Muse Spark Open evaluation methodology (PDF)
• Meta Model API docs: models
• Apple Machine Learning Research: MLX distributed inference
• Apple: Mac Studio
• Artificial Analysis: Muse Spark Open
• Artificial Analysis: open-weights leaderboard
• Artificial Analysis: Muse Spark Open vs Kimi K3
• VentureBeat: Meta unveils Muse Spark Open

*Not sure whether an open model you host yourself beats one you rent? Take the free 2-minute quiz and get matched. Then read the closed model Meta shipped three days earlier, the one that actually scores in the frontier tier: Muse Wonder.*