Meta released the weights of Muse Spark Open on Oct. 152. Hugging Face published the recipe for running it the same day1. The model holds about 816 billion parameters and uses about 42 billion for each token1. It reads a million tokens of text, images, audio and video1. Meta calls it "the most capable US open-weight model to date"2. That claim comes with a license, a hardware bill and a benchmark table Meta wrote itself3. The AILately.com Desk read all three.
What Meta Shipped
Muse Spark Open is a mixture-of-experts model. Each of its 62 layers holds 256 small expert networks, and a router picks 8 of them for every token1. Only those 8 run, so the model computes at the speed of a 42-billion-parameter model1. All 256 stay in memory, so it needs the memory of an 816-billion-parameter one1. Its full-precision checkpoint is about 1.5 terabytes1. A 4-bit version for Nvidia's Blackwell chips fits on one B200 server1.
Glimmer came first. Meta released that 30-billion-parameter model, Muse Glimmer, in August under the Apache 2.0 license7. Hugging Face counts 1,016,198 downloads of it7. Muse Spark Open uses the same tokenizer and chat template1. Reasoning stays on at all times1. A developer can set it to low, medium, high or xhigh, and the default is xhigh1.
Hugging Face's engineers fine-tuned the model on 64 H100 chips for 500 steps, which took about two hours1. They also ran a 2-bit version on one Mac Studio with 512 GB of memory at 18 tokens a second1.
The Mac Pitch
Meta built its launch around Apple. "Muse Spark Open is the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework," the company wrote2. Each Mac holds up to 512 GB of unified memory, so two of them hold the whole model2. Mark Zuckerberg posted the same message on X. "Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple," he wrote, as The New Stack quoted him4.
Frederic Lardinois of The New Stack priced that desktop. "Apple doesn't currently sell the 512 GB Mac Studio to mere mortals, but in a few weeks, you will be able to buy two of those machines for the price of a compact car," he wrote4. Lucas Ropek of TechCrunch put the memory need at a terabyte, citing Engadget5. Meta's model card asks for 700 GB of graphics memory for the 4-bit version and 2 TB for full precision3.
The Benchmarks, Read Both Ways
Meta's model card compares Muse Spark Open with seven open-weight rivals at its xhigh setting3. Against the two American models, Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra, it leads every row34. JobBench, a test of agent work, shows the gap3. Muse Spark Open scored 61.8, GLM 5.3 scored 61.4 and Kimi K3 scored 54.3, while Inkling scored 31.1 and Nemotron 3 Ultra 27.2. MCP-Atlas, a tool-use test, puts Muse Spark Open at 89.2 against 76.0 for Inkling3.
Muse Spark Open scores 61.8 on JobBench against 31.1 for Inkling, the closest American rival
JobBench scores for five open-weight models, from Meta's model card; higher is better
Each bar is one open-weight model's score on JobBench, a test of agent work; the magenta bar is Muse Spark Open and the gray bars are the two American rivals Meta names.
Source: Hugging Face [3]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value |
|---|---|
| Muse Spark Open | 61.8 |
| GLM 5.3 | 61.4 |
| Kimi K3 | 54.3 |
| Inkling | 31.1 |
| Nemotron 3 Ultra | 27.2 |
Chinese models tell a different story. Terminal-Bench 2.1, a coding test, gave Muse Spark Open 76.4, Kimi K3 85.0 and GLM 5.3 83.93. DeepSWE 1.1, another coding test, gave Kimi K3 68 and Muse Spark Open 603. Meta's own card lists the Artificial Analysis Intelligence Index at 39 for its model, 45 for GLM 5.3 and 44 for Kimi K33. Lardinois read the coding gap against Meta's April launch of the closed Muse Spark. "That's the same gap Meta acknowledged when it launched the closed version of Spark in April, and it doesn't look like it's made much progress since," he wrote4.
Muse Spark Open trails Kimi K3 by 8.6 points on Terminal-Bench 2.1 and by 8 on DeepSWE 1.1
Scores on two coding benchmarks, percent of tasks solved, from Meta's model card; higher is better
Each row is one coding benchmark: the blue dot is Muse Spark Open and the magenta dot is Kimi K3, the open-weight model from Moonshot AI that Meta's own table puts ahead.
Source: Hugging Face [3]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Muse Spark Open | Kimi K3 |
|---|---|---|
| Terminal-Bench 2.1 | 76.4% | 85.0% |
| DeepSWE 1.1 | 60.0% | 68.0% |
Muse Spark Open wins where tools and images meet. OSWorld Verified, a computer-use test, gave it 80.4 against 84.8 for Kimi K3 and 40.1 for Inkling3. ZeroBench with tools gave it 50.0 against 41.0 for Kimi K33. HLE-Diamond, taken with tools switched off, gave it 26.1 against 22.23.
Artificial Analysis, an independent evaluator, ran its own index. Muse Spark Open scored 39, tenth of 120 models, against a median of 196. The same page calls it "particularly expensive when comparing to other open weight models of similar size"6. Hosted providers charge $1.25 per million input tokens and $4.25 per million output tokens6. It generated 118 tokens a second6.
The License
Muse Spark Open's weights ship under the Muse Model License3. It opens with the words of the MIT license and then adds two conditions3. Any product that uses the model, or any model trained or distilled from its outputs, must "prominently display 'Built with Muse Spark Open'" in its interface or documentation3. A company that sells model access as a service and passes $20 million in revenue over any 12 months "must notify Meta and enter into a separate written agreement with Meta" first3. Meta leaves ownership of outputs with the user3.
Its usage policy bars military, nuclear and espionage uses3. It also bars the operation of critical infrastructure, transportation and heavy machinery, and any use by people under 183. A download starts with a form, since the repository is gated3.
Lardinois placed the license beside its rivals. "Overall, this looks a lot like the license that Moonshot uses for Kimi K3, for example," he wrote4. Meta's own post describes the terms as "limited commercial terms, which allow many developers to use the model for free"2.
The Safety Table
Meta's card carries numbers it rarely prints beside rivals. StrongREJECT v2 is an adaptive jailbreak test where lower is better3. Muse Spark Open scored 0.5 on it, against 66.6 for Kimi K3, 62.9 for GLM 5.3 and 86.4 for Inkling3. Meta says the open version carries "additional safety post-training to make it more robust against misuse as a standalone model without system guardrails"3. It refused 99.2% of jailbreak attempts on Meta's internal cyber benchmark, against 75.5% for Kimi K3 in the same setup3.
Muse Spark Open scored 0.5 on the StrongREJECT jailbreak test against 66.6 for Kimi K3
StrongREJECT v2 adaptive jailbreak score from Meta's model card, where lower is better
Each bar is one model's score on an adaptive jailbreak test from Meta's own card; a shorter bar means the model gave in to fewer attacks, and the magenta bar is Muse Spark Open.
Source: Hugging Face [3]. Chart by The AILately.com Desk.
The numbers behind this chart
| Item | Value |
|---|---|
| Muse Spark Open | 0.5 |
| GLM 5.3 | 62.9 |
| Kimi K3 | 66.6 |
| Inkling | 86.4 |
The Case For and the Case Against
Meta's case is reach. "We believe the future is for everyone, and that includes the freedom to run advanced models wherever, including at home," the company wrote2. Zuckerberg's post names the partner behind the home claim, Apple, and the machine, a Mac Studio4. Seven hosting companies served the model on launch day, and the weights sit on four clouds2.
Lardinois makes the case against the word open. "Like with some of the newer Chinese models, though, the weights may be open but the license doesn't tell the same story," he wrote4. His headline called the license "semi-open"4. The first independent score agrees with Meta on rank and parts with it on price6. It places the model tenth of 120 open models and calls it expensive for its size.
What Each Side Needs
Meta's case needs the full evaluation report it promised2. Every benchmark row on the card came from Meta's own harness3. It also needs the 512 GB Mac Studio on sale. Lardinois's case needs a developer to hit the $20 million clause or balk at the credit line4.
What to watch
- Meta's full evaluation report, which the card says is coming2.
- Audio and video scores from Artificial Analysis, whose page carries only text results so far6.
- Hugging Face download counts, which stood at 815 for the month on Oct. 16 against 1,016,198 all time for Muse Glimmer37.
- A Terminal-Bench run of Muse Spark Open outside Meta's harness, against Kimi K3's 85.03.
Sources
- Pedro Cuenca, Merve Noyan, Ben Burtenshaw, Guarin Flück, Quentin Gallouédec, Alvaro Bartolome, Serveurperso and Jun Kim, "Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta," Hugging Face, Oct. 15, 2026, https://huggingface.co/blog/meta-muse-spark-open
- Meta, "Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date," Meta AI Research, Oct. 15, 2026, https://research.meta.ai/blog/muse-spark-open
- Meta Superintelligence Labs, "Muse Spark Open Model Card (meta-models/Muse-Spark-Open-816B-A42B)," Hugging Face, Oct. 15, 2026, https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B
- Frederic Lardinois, "Meta ships Muse Spark Open, its answer to China's best open-weight models," The New Stack, Oct. 15, 2026, https://thenewstack.io/meta-muse-spark-open/
- Lucas Ropek, "Meta launches an open-weight AI model that runs on Macs," TechCrunch, Oct. 15, 2026, https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/
- Artificial Analysis, "Muse Spark Open (xhigh) - Intelligence, Performance & Price Analysis," Artificial Analysis, Oct. 16, 2026, https://artificialanalysis.ai/models/muse-spark-open
- Meta, "Muse Glimmer 30B Model Card (meta-models/Muse-Glimmer-30B)," Hugging Face, Aug. 11, 2026, https://huggingface.co/meta-models/Muse-Glimmer-30B