By Thorsten Meyer | October 16, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta finally put a big Muse model on Hugging Face. It scores 39, hallucinates less than any top open model, and costs the same per token as the closed Muse Spark 1.3, which scores higher.
On October 15, 2026, Meta released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token. The weights are on Hugging Face. Artificial Analysis tested it the same day at its default xhigh reasoning effort.
The headline number: 38.9 on the Artificial Analysis Intelligence Index v4.3.2, which Artificial Analysis rounds to 39. That makes it the highest-scoring open-weight model from a US lab by a wide margin. It is also the tenth-best open-weight model overall. The nine ahead of it all come from Chinese labs.
That’s the strategic story. The practical story is the license and the bill, and you need to understand both before you build on it.
What Meta shipped
- Size: 816B total, 42B active parameters, 256 experts with 8 active per token
- Context window: 1 million tokens
- Input: text, images, speech/audio and video. Output: text
- Weights: BF16 and NVFP4 on Hugging Face. Meta’s model card puts self-hosting at 700 GB of GPU memory for NVFP4 and 2 TB for BF16, and says the model is optimized to run on two Mac Studios with M5 Ultra using MLX
- API: four third-party providers, with a median list price of $1.25 per million input tokens, $0.15 for cached input and $4.25 per million output tokens
That is exactly the list price of Meta’s proprietary Muse Spark 1.3, cache price included. The weights are free. Rented by the token, you pay closed-model rates.
high performance open-weight AI model
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Where it lands
Muse Spark Open is the first open-weight model from Meta’s flagship Muse Spark line. Meta’s recent open models were much smaller. Muse Glimmer (30B, Apache 2.0, August 2026) scores 17 on the current index. Llama 4 Maverick (April 2025) scores 9.
Infographic 01 · Where Muse Spark Open lands
The best US open-weight model is tenth. The nine ahead are Chinese.
Artificial Analysis scores Meta’s first open Muse Spark model at 38.9. That is 12 points clear of any other US open-weight model, but behind nine Chinese models. Four of those nine are MIT-licensed, and three of them cost a third or less per task.
Top open-weight models by Intelligence Index
Higher is better. Effort variants of the same model are not listed twice. Cost = average API cost to run one index task.
License: Muse Spark Open ships under the Muse Model License: broad use and redistribution, but mandatory “Built with Muse Spark Open” attribution, a separate agreement for model-as-a-service businesses above $20M revenue, and a usage policy.
Source: Artificial Analysis, Muse Spark Open (xhigh) model page and model data, read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations. License categories as listed by Artificial Analysis. Countries: Xiaomi, Z AI, StepFun, Kimi, InclusionAI, Alibaba and DeepSeek are Chinese labs; Meta and NVIDIA are US.
Powered by Thorsten Meyer AI
Among US open-weight models, nothing else comes close. NVIDIA’s Nemotron 3.5 Super VL, released the same day, scores 26.5. Artificial Analysis calls it the leading US open-weight model, and among permissively licensed US models it is. Muse Spark Open scores 12 points higher, but under a license with conditions. I cover the NVIDIA release in a separate piece.
Against the whole open-weight field, the picture is less flattering. MiMo-V2.6-Pro (Xiaomi, MIT license) leads at 46.3. GLM-5.3, Step 5 Preview, Kimi K3, GLM-5.3-Flash, Ling 3.1 Flash, two Qwen3.8 models and DeepSeek V4.1 Flash all score higher too. My summer look at open models described Chinese labs setting the size ceiling while US open releases stayed small. An 816B model from Meta changes the size story. It doesn’t change who leads on score.
Meta’s own family is the more uncomfortable comparison. At the same xhigh effort and the same list price, the closed Muse Spark 1.3 scores 45.1, 6 points more. Muse Wonder, which I covered earlier this week, scores 52.4. The open model lands about where Muse Spark 1.2 (xhigh) did in August: 39.6.
As an affiliate, we earn on qualifying purchases.
Where it is genuinely strong
- AA-Omniscience: 24.7, first among all 167 open-weight entries. This index rewards correct answers and penalizes confident wrong ones. The next-best open model, Kimi K3, scores 19.7. MiMo-V2.6-Pro scores 8.4, and DeepSeek V4.1 Flash scores −5.3.
- Hallucination rate: 24%, the lowest among the 15 highest-scoring open-weight models. GLM-5.3-Flash is at 28%, MiMo-V2.6-Pro at 41% and Kimi K3 at 53%.
- MMMU-Pro (visual reasoning): 81%, first among the 66 open-weight entries tested.
- GPQA Diamond 89%, Humanity’s Last Exam 44% (fourth among open weights), SciCode 58% (fifth), AA-LCR long-context reasoning 78%.
A low hallucination rate, strong visual reasoning and a 1M context window make a coherent package for document and research work.
As an affiliate, we earn on qualifying purchases.
Where it is not: agents in a terminal
Meta’s model card says Muse Spark Open is “built for coding and agentic work”. Meta’s own figures support that on some tests: 89.2 on MCP-Atlas, 80.4 on OSWorld Verified and 76.4% on Terminal-Bench 2.1, a number Artificial Analysis matches at 76%.
On Terminal-Bench 4.0, the harder version in the Intelligence Index, Artificial Analysis measures 8%. Muse Spark 1.3 scores 33% there, MiMo-V2.6-Pro 35% and GLM-5.3 42%. The other agentic results point the same way: AutomationBench-AA 50% (DeepSeek V4.1 Flash: 69%), GDPval-AA 1401 Elo (MiMo-V2.6-Pro: 1684, Muse Spark 1.3: 1681) and AA-Briefcase 1279 Elo (Muse Spark 1.3: 1581).
The 8% also shows up in the bill: Terminal-Bench 4.0 accounts for $0.47 of the $0.84 weighted cost per index task. Long, failing agent runs are expensive.
I would not run unattended agents in a shell on this model. The Chinese open leaders are clearly ahead there, and my reasons for not running agents on Mistral Large 4 apply here too.
As an affiliate, we earn on qualifying purchases.
The price of a 39
Infographic 02 · Profile and price
Reliable on knowledge, weak in the terminal, priced like the closed model
Muse Spark Open matches its closed sibling on knowledge reliability and leads the open-weight leader there by a wide margin. On Terminal-Bench 4.0 it scores 8%. It reads fewer input tokens than its MIT-licensed rivals and still pays the most for them.
The ten Intelligence Index evaluations
Higher is better. Bars use Artificial Analysis’s normalisation (Elo: (score − 500) / 2000; AA-Omniscience: negative values shown as empty). Labels show raw values.
- AA-BriefcaseElo127915811516
- GDPval-AAElo140116811684
- AutomationBench-AAscore50%58%59%
- Terminal-Bench 4.0score8%33%35%
- SciCodescore58%59%61%
- Humanity’s Last Examscore44%49%49%
- GDP.pdfscore19%27%19%
- CritPtscore15%25%27%
- AA-Omniscienceindex, −100 to 10024.725.08.4
- AA-LCRscore78%83%86%
Two Terminal-Benches: on the older Terminal-Bench 2.1, Muse Spark Open scores 76%, close to Meta’s own 76.4%. Terminal-Bench 4.0 is the harder version used in the index, and it accounts for $0.47 of the model’s $0.84 weighted cost per task.
Where the $0.84 goes
Per Intelligence Index task at xhigh effort
List price $1.25 input, $0.15 cached input, $4.25 output per 1M tokens: identical to the proprietary Muse Spark 1.3.
Reads less, pays more
Input tokens over the full index run vs. what they cost
| Model | Input tokens | Cached / 1M | Input $/task | Total $/task |
|---|---|---|---|---|
| Muse Spark Open | 2.96B | $0.15 | $0.67 | $0.84 |
| MiMo-V2.6-Pro | 4.43B | $0.0036 | $0.08 | $0.13 |
| GLM-5.3-Flash | 6.48B | $0.03 | $0.25 | $0.28 |
| DeepSeek V4.1 Flash | 9.26B | $0.006 | $0.16 | $0.27 |
Muse Spark Open uses the fewest input tokens of the four, but its cached-token price is roughly 40× MiMo-V2.6-Pro’s.
Show the evaluation data as a table
| Evaluation | Muse Spark Open | Muse Spark 1.3 | MiMo-V2.6-Pro |
|---|---|---|---|
| AA-Briefcase | 1279 | 1581 | 1516 |
| GDPval-AA | 1401 | 1681 | 1684 |
| AutomationBench-AA | 50% | 58% | 59% |
| Terminal-Bench 4.0 | 8% | 33% | 35% |
| SciCode | 58% | 59% | 61% |
| Humanity’s Last Exam | 44% | 49% | 49% |
| GDP.pdf | 19% | 27% | 19% |
| CritPt | 15% | 25% | 27% |
| AA-Omniscience | 24.7 | 25.0 | 8.4 |
| AA-LCR | 78% | 83% | 86% |
| Intelligence Index | 38.9 | 48.1 | 46.3 |
Source: Artificial Analysis, Muse Spark Open (xhigh) model page and model data, read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations. License categories as listed by Artificial Analysis. Muse Spark 1.3 at max effort (proprietary); MiMo-V2.6-Pro by Xiaomi (MIT).
Powered by Thorsten Meyer AI
One Intelligence Index task costs $0.84 on Muse Spark Open. Here is what the open-weight neighbours charge:
- MiMo-V2.6-Pro (MIT): 46.3 for $0.13
- MiMo-V2.6-Flash (MIT): 37.9 for $0.06
- DeepSeek V4.1 Flash (MIT): 39.5 for $0.27
- GLM-5.3-Flash (MIT): 41.8 for $0.28
- Step 5 Preview (own license): 43.7 for $1.03
- Kimi K3 (own license): 43.6 for $2.00
- GLM-5.3 (own license): 44.8 for $2.01
MiMo-V2.6-Pro scores 7 points more for about a sixth of the cost; DeepSeek V4.1 Flash matches the score for a third. Among models with their own restrictive licenses, Muse Spark Open is actually cheaper than Step 5 Preview, Kimi K3 and GLM-5.3. Against MIT-licensed models, it isn’t competitive on price.
Again, the bill is in the input
For Muse Wonder, I wrote that Meta’s bill is in the input. The same holds here, for a different reason.
Of the $0.84, $0.67 (80%) is input: $0.40 in cache reads, $0.25 in cache writes and $0.02 in uncached input. Output is $0.17.
Muse Spark Open doesn’t read an unusual amount. Across the full index run it consumed 2.96 billion input tokens, fewer than MiMo-V2.6-Pro (4.43B), GLM-5.3-Flash (6.48B) or DeepSeek V4.1 Flash (9.26B). It also writes less: 40,000 output tokens per task, the lowest among the ten top-scoring open models. Kimi K3 writes 48,000 and DeepSeek V4.1 Flash 89,000.
The difference is the price of a cached token: $0.15 per million on Muse Spark Open, $0.0036 on MiMo-V2.6-Pro (roughly 40 times less), $0.006 on DeepSeek V4.1 Flash and $0.03 on GLM-5.3-Flash. An 88% cache discount sounds generous until you compare the result with the open-weight market. In agent loops that re-read the same context, the cache price matters more than the headline input price.
Self-hosting avoids the per-token rate, if you have 700 GB of GPU memory to spare.
Speed
The four providers deliver a median 102 output tokens per second (126 t/s with 100k-token prompts). The first chunk arrives after 0.67 seconds, but that is the reasoning stream; the first answer token takes about 20 seconds at xhigh. Fine for batch work. For interactive use, lower the effort.
The license: read it before you build
Artificial Analysis classifies Muse Spark Open as open weights with restricted commercial use. Its Openness Index gives the model 28 out of 100, with zero points for both pre-training and post-training data transparency. MiMo-V2.6-Pro scores 53.
Downloading requires accepting Meta’s terms on Hugging Face and submitting your full legal name, date of birth and organization. I read the Muse Model License as shown on the model page. These are the parts that matter most for builders:
- The grant is broad. It allows use, modification, fine-tuning, distillation, redistribution and sale. Meta claims no ownership of outputs.
- Attribution is mandatory. Products using the model or any derivative must prominently display “Built with Muse Spark Open”. The license defines derivatives to include models trained on its outputs.
- There’s a revenue threshold for model hosting. If you offer models as a service through an API and your group’s revenue exceeds $20 million in any 12 months, you must notify Meta and sign a separate agreement before offering Muse Spark Open that way. Internal use, research and using the model inside your own product are exempt.
- There is a usage policy. Among other things, it rules out military, warfare, nuclear and espionage uses, operating critical infrastructure and the unlicensed practice of professions such as law or medicine. It also says the model is not intended for people under 18.
For most companies building a product, that’s workable. For large inference providers, defence-adjacent work and legal teams that insist on MIT or Apache 2.0, it isn’t. My summary is not legal advice; have your own counsel read it.
Who should use it
If you need a US-origin open-weight model for procurement, compliance or policy reasons, Muse Spark Open is now the strongest option by far.
If you do knowledge-heavy work on documents, images or long contexts where hallucinations are expensive, its AA-Omniscience lead and 24% hallucination rate make it worth a real evaluation on your own data.
If you want cheap open-weight intelligence through an API, this is the wrong model. MiMo-V2.6-Pro, DeepSeek V4.1 Flash and GLM-5.3-Flash give you the same or better scores for a third to a sixth of the cost, under MIT.
If you run agents in a terminal, skip it. 8% on Terminal-Bench 4.0 is a clear signal.
None of this changes my October 2026 AI stack. An open model would have to beat the Chinese open-weight leaders on price or agent work to get a slot there. Muse Spark Open does neither.
Verdict
Muse Spark Open matters more as a signal than as a product. A US frontier lab has released an 816B open-weight model that beats every other US open model by 12 points and leads all open models on knowledge reliability and visual reasoning. For anyone worried that open-weight frontier work had become a Chinese monopoly, that is good news.
As a product, it is harder to defend. It scores 6 points below the closed Muse Spark 1.3 at the same effort and list price, struggles on agentic terminal work and costs three to six times as much per task as MIT-licensed models that score the same or better. And it comes with a license you have to read.
Meta opened the weights. It did not open the price.
Sources
- Artificial Analysis: Muse Spark Open (Xhigh) model page (Intelligence Index v4.3.2, read October 16, 2026)
- Artificial Analysis: NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt (October 15, 2026)
- Artificial Analysis model data for all compared open-weight and Meta models, read October 16, 2026 (scores, cost per index task, token counts, prices, speed)
- Meta: Muse Spark Open model card and Muse Model License on Hugging Face (October 15, 2026)
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
