Meta released the weights of Muse Spark Open on October 15, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters per token and a 1M-token context, built for coding and agent work. It takes text, image, video and audio as input and returns text. Meta’s announcement PyTorch blog Meta calls it “the most capable US open-weights model to date”; that is Meta’s own ranking, not an outside one. PyTorch blog The decision it puts in front of a team is whether to pay $1.25 per million input tokens and $4.25 per million output tokens at a hosted provider, or to find at least 700 GB of GPU memory, or two 512 GB Mac Studios, and run it under the Muse Model License. Fireworks Model card
S5 Labs has not run the model. Every benchmark, throughput and memory figure below belongs to Meta, Apple or the provider named.
Which Muse this is
Meta now has several models named Muse, and the three that matter here are different products. Muse Spark Open “shares the same base model as Muse Spark 1.3”, according to the PyTorch post, which also says the open version received additional safety post-training so it can run without the system guardrails Meta wraps around its API. PyTorch blog Spark 1.3 is the model Meta put behind Muse Code and the Meta Model API on September 2. Our Spark 1.3 coverage Muse Glimmer, which the PyTorch post describes only as “a small dense model”, is the 30B release from August that we covered for fitting a 24 GB card under Apache 2.0. Muse Spark Open does not use Apache 2.0.
| Item | Muse Spark Open (model card) |
|---|---|
| Parameters | 816B total, 42B active per token |
| Architecture | MoE transformer, 62 layers, hidden size 8192, 256 experts with 8 active |
| Attention | 128 query heads, 16 KV heads, head dim 64 |
| Vocabulary | 202,048 |
| Vision/audio | 1.8B Perception Encoder, 50 layers |
| Modalities | Text, image, video, audio in; text out |
| Context | 1M tokens |
| Weights | BF16 and NVFP4 |
| Sampling | temperature 1.0, top_p 1.0, top_k disabled |
| reasoning_effort | low, medium, high or xhigh; default xhigh, anything else errors |
Source: model card.
The weights are gated. The Hugging Face form asks for a full legal name, date of birth and full organization name with corporate identifiers written out, and submissions cannot be edited after the fact. Model card
Weight size is not runtime memory
Apple’s post puts the NVFP4 parameters at about 493 GB and BF16 at more than 1.6 TB for parameters alone. Apple, MLX distributed inference The model card’s requirements are higher: at least 700 GB of VRAM for NVFP4 and at least 2 TB for BF16. Model card The card does not break down what fills the gap between 493 GB of weights and the 700 GB floor. Treat 700 GB as the floor for the 4-bit build and budget the 1M-token context separately.
The card also carries an unusual line: the model is optimized to run on two Mac Studios with M5 Ultra using MLX. Model card Meta’s blog explains the arithmetic. Each Mac Studio’s GPU can address up to 512 GB of unified memory with 1.2 TB/s of bandwidth, and two of them together provide about a terabyte of unified memory, which Meta says is enough to hold the full model. Meta’s announcement Apple’s post says the model does not fit on a single 512 GB Mac Studio, even though the weights alone come in under that figure. Apple, MLX distributed inference We have not checked Apple’s current configurations or pricing for a 512 GB M5 Ultra.
What two Mac Studios deliver, per Apple
The only published throughput for the Mac setup comes from Apple, not Meta. Apple evaluated the NVFP4 build over Thunderbolt 5 with RDMA, which needs MLX 0.30.1 or later for the jaccl backend and macOS 26.2 or later for the driver. Apple, MLX distributed inference
| Setup (Apple’s runs, NVFP4) | Decode tok/s, 8,192 | Decode tok/s, 65,536 | TTFT, 8,192 | TTFT, 65,536 |
|---|---|---|---|---|
| 2 × M5 Ultra, pipeline parallel | 15.74 | 14.37 | 8.4 s | 47.6 s |
| 2 × M5 Ultra, tensor parallel | 24.99 | 23.84 | 8.2 s | 70.3 s |
| 4 × M5 Ultra, tensor parallel | 36.73 | 34.66 | 4.8 s | 40.6 s |
Source: Apple, Tables 2 and 3. Apple’s tables have one tab per context size, 8,192 and 65,536 tokens, and the figures above are read from each tab. Decode speed was measured over 128 generated tokens after the context.
Longer context costs little decode speed but a lot of waiting. Decode drops about 5 to 9 percent going from 8,192 to 65,536 tokens, while time to first token grows roughly six-fold on two machines with pipeline parallelism (8.4 to 47.6 seconds) and roughly nine-fold with tensor parallelism (8.2 to 70.3 seconds). At 65,536 tokens, two-machine tensor parallelism loses to two-machine pipeline parallelism on prefill because of communication overhead. For the four-machine tensor setup, Apple says the 40.6-second figure is with two Thunderbolt cables per link, and that one cable takes about 50.5 seconds. Apple, MLX distributed inference An agent loop that fills 65K of context before each turn will feel that wait.
Apple’s launch command is a single line:
mlx.launch --hostfile hostfile.json --backend jaccl --env MLX_METAL_FAST_SYNCH=1 -- mlx_lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080Apple, MLX distributed inference
For scale: at Apple’s 24.99 tokens per second (two Macs, tensor parallel, 8,192-token context), a two-Mac rig generating without pause for 24 hours would produce about 2.16 million output tokens, which Fireworks or Together would bill at roughly $9 at their listed output price. That is one request stream, and at 65,536 tokens Apple’s 23.84 tokens per second works out to about $8.75 a day, so it flatters neither side. It does show that the self-hosting case rests on data control, the license, or keeping prompts off a third party’s servers rather than on cost per token.
Renting it
Meta lists weights on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models, and API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta’s announcement Supported engines are SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth; the PyTorch post says the SGLang and vLLM paths are tuned for NVIDIA, AMD and Intel hardware. Meta’s announcement PyTorch blog
Two providers had prices up when we checked, and they match: $1.25 per million input tokens, $0.15 for cached input and $4.25 per million output. Together Fireworks Fireworks describes it as serverless on US-only capacity with no US-only premium and a 1040K context; the page does not state the serving precision, though its fine-tuning base is named muse-spark-open-nvfp4. Fireworks Together’s page is contradictory: its spec text says “Available on Together AI” with serverless deployment while a banner says “Launching soon” and asks for an email. Together We have not confirmed prices at Baseten, Groq, Modal, Nebius or OpenRouter.
The license
The model card’s license summary allows use, modification including fine-tuning, distillation and redistribution, subject to conditions. Model card The two conditions that bite:
- Attribution. “Built with Muse Spark Open” must be displayed prominently.
- Model-as-a-service threshold. If you or your affiliates offer access to the model to third parties through an API or something similar, and aggregate gross revenue from that exceeds $20 million in any consecutive 12 months, you must notify Meta and sign a separate agreement before offering it.
The exclusions matter as much as the threshold. Internal use, research, commercial activity that is not model-as-a-service, and use of the model as a component inside your own product are all outside the $20 million clause. Model card Read that way, the clause targets inference resellers, not companies that ship an app with the model inside it. Meta claims no ownership of outputs, the software is provided “as is”, an Exhibit A usage policy applies, and the model is not intended for users under 18. Model card
We read the card’s summary, not the full license text, so the termination, governing-law and naming provisions are unknown to us, and the Exhibit A list is unverified here. A team that plans to redistribute or fine-tune should read the whole document before an evaluation becomes a dependency.
Meta’s benchmark table
Meta’s comparison runs its own model at xhigh reasoning effort, Kimi K3 and GLM 5.3 at max, and Qwen3.8-2.4T-A95B at xhigh. Model card A selection of the rows:
| Benchmark (Meta’s runs) | Muse Spark Open | Kimi K3 | GLM 5.3 | Qwen3.8-2.4T-A95B |
|---|---|---|---|---|
| AA Intelligence Index | 39 | 44 | 45 | 40 |
| GPQA Diamond | 89.0 | 93.5 | 91.7 | 93.5 |
| HLE (with tools) | 61.0 | 56.0 | 62.5 | 56.2 |
| HLE-Diamond (no tools) | 26.1 | 22.2 | 16.4 | 17.3 |
| MCP-Atlas | 89.2 | 82.3 | 84.2 | 84.5 |
| Toolathlon Verified | 72.8 | 76.5 | 73.0 | 72.5 |
| Terminal-Bench 2.1 (Terminus 2) | 76.4 | 85.0 | 83.9 | 82.0 |
| DeepSWE 1.1 (preferred harness) | 60 | 68 | 67 | 57 |
| SWE-Atlas-QnA (preferred harness) | 58 | 66 | 59 | 62 |
| SciCode | 58.5 | 59.5 | 59.0 | 54.1 |
| OSWorld Verified | 80.4 | 84.8 | not reported | not reported |
| JobBench | 61.8 | 54.3 | 61.4 | 52.7 |
| GDPVal-AA v2 | 1401 | 1537 | 1653 | 1613 |
| AA Omniscience | 25 | 20 | 14 | 4 |
| StrongREJECT v2 (adaptive, lower is better) | 0.5 | 66.6 | 62.9 | 65.6 |
Source: model card.
On the tool-use rows Meta chose, MCP-Atlas, HLE with tools, JobBench and the two vision rows the card reports against Kimi only (ZeroBench 50.0 vs 41.0, BabyVision 80.2 vs 85.7), Muse Spark Open leads or trades blows. Model card On the coding rows it does not: Kimi K3 is 8.6 points ahead on Terminal-Bench 2.1 and 8 points ahead on DeepSWE 1.1, and GLM 5.3 beats it on both. The card’s own extra columns make the same point. Reflection’s Beam scores 80.1 on Terminal-Bench 2.1 against Meta’s 76.4, and Mistral Large 4 scores 62 on DeepSWE 1.1 and 1424 on GDPVal-AA v2, both above Muse Spark Open. Model card The StrongREJECT row, where Meta’s model refuses adaptive jailbreaks at a rate the others do not approach, is presumably what the additional safety post-training in the PyTorch post bought. The card does not measure whether that also shows up as over-refusal on legitimate work.
Kimi K3 is the comparison most teams will care about, since it is the open-weight model the card measures against most often. Its weights are 2.8T total and take about 1.4 TB of fast memory to run, which puts a two-Mac-Studio setup out of reach. Beam, at 501B total under Apache 2.0, is the closer match on license and size. We have not opened Meta’s evaluation methodology report, so we cannot say whether the temperature 1.0 and top_p 1.0 defaults were also the benchmark settings.
What to settle before committing
- If you are choosing between renting and hosting, count the data-control and license reasons first. At the listed hosted prices and Apple’s throughput, cost per token does not favor two Mac Studios.
- If you plan to host on GPUs, start from 700 GB for NVFP4, not from Apple’s 493 GB, and ask what context length that figure assumes. The card does not say.
- If the Mac route is the point, confirm the 512 GB M5 Ultra configuration and its lead time yourself, and test at the context your agent reaches after twenty turns. Apple publishes nothing past 65K, and even there its best first-token time is about 40 seconds, on four machines with two cables per link.
- If you resell inference, model the $20 million threshold across affiliates. If you ship the model inside a product, the summary says the clause does not apply, but read the full license, which we have not.
- Compare against Kimi K3 and Beam on your own coding harness. Meta’s table puts both ahead of Muse Spark Open on Terminal-Bench 2.1.
Our open models guide tracks the license and size trade-offs across these releases, and the local LLM machines guide covers what a 700 GB memory floor means in hardware.