Muse Spark Open: 816B, 1M context, $1.25 out

On October 15 Meta Superintelligence Labs released the weights for Muse Spark Open, a mixture-of-experts model with 816B total parameters and 42B active per token, a 1M-token context window, and native multimodal input covering image, video, audio and PDF. Meta’s own blog calls it “the most capable US open-weight model to date.” The weights went up on Hugging Face the same day, and the safetensors metadata records a total file size of 1,644,333,584,632 bytes — about 1.64TB.

Key points

How big the weights are, and what it takes to run them. Apple’s MLX engineering post gives the two numbers that matter: at native BF16 the weights need “more than 1.6TB” of memory, and quantized to NVFP4 4-bit the footprint is “about 493GB.” A single Mac Studio tops out at 512GB of unified memory, so Meta’s minimum local demo is two M5 Ultra Mac Studios linked over Thunderbolt 5 RDMA.
Measured throughput, not a chart. The same post reports decode speed at an 8,192-token context over 128 generated tokens: 24.99 tokens/s on two machines with tensor parallelism, 15.74 tok/s with pipeline parallelism, and 36.73 tok/s across four machines. At 65,536 tokens of context, running two Thunderbolt cables per link cut time to first token from about 50.5 seconds to 40.6 seconds on four machines.
A real price. OpenRouter’s official post at 18:02 on October 15 reads “$1.25/M in, $4.25/M out,” with the model live day 0 alongside Baseten, Fireworks, Nebius and Together. Meta’s blog lists seven API platforms in total.

Background

The Muse Spark line took three steps to get here. Meta launched Muse Spark in August alongside its Muse Code coding tool — the PyTorch blog links directly to that release — updated it to Spark 1.3 roughly six weeks ago, and on October 15 opened the weights of the 1.3 base model. The PyTorch post says it “shares the same base model as Muse Spark 1.3” and notes this version carries extra safety post-training for the case of the model “being used as a standalone model without system guardrails.” Three days earlier, on October 12, the same lab shipped the paid frontier models Muse Wonder and Muse Wonder Full at $4/$20 per million tokens, with the strongest variant locked to a subscription. This release is an order of magnitude cheaper and downloads. The second thread is Meta and Apple, usually rivals: The Verge reports Zuckerberg posting on X that the model “runs great on new Mac Studios thanks to work with @johnternus and Apple,” with Ternus replying that it is “amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk.” Axios reported the two companies had been working together for months.

The facts

Specs and the self-hosting bar (Hugging Face model card and the Apple MLX post, both read on October 16):

Item	Number	Whose figure
Total / active parameters	816B / 42B	Meta blog, PyTorch blog, HF summary all agree
safetensors metadata	822,036,529,664 parameters	HF model card; the gap is non-language modules
Total weight files	1,644,333,584,632 bytes (~1.64TB)	HF model card `totalFileSize`, 65 shards
BF16 self-hosting memory	More than 1.6TB (HF says 2TB)	Apple MLX post; HF self-hosting row
NVFP4 self-hosting memory	About 493GB; HF says a GPU with 700GB+ VRAM	Apple MLX post; HF self-hosting row
Context	1M tokens	Meta blog, OpenRouter post
Decode speed (8,192 context)	2 machines tensor 24.99, pipeline 15.74, 4 machines tensor 36.73 tok/s	Apple MLX post, Table 3
API price	$1.25/M input, $4.25/M output	OpenRouter official post

The price needs a flag. The Hugging Face card’s body says the model “is available via API through third party inference providers,” while Artificial Analysis’s page shows price “N/A” and speed “Unknown” and states there is no publicly available API to measure. OpenRouter’s own post is the primary figure here; the AA and HF fields look like day-one data that had not landed yet.

The license is worth spelling out. The Muse Model License permits free use, modification, fine-tuning, distillation, merging, distribution and sale, with three conditions: ship the copyright notice, license and Usage Policy with copies and derivatives; display “Built with Muse Spark Open” in the UI of any product or service using it, or in its docs, website or settings; and if the licensee or its affiliates offer model-as-a-service to third parties with aggregate gross revenue above $20 million in any consecutive 12 months, notify Meta and sign a separate written agreement first. Meta “claims no ownership of outputs,” and the model ships as is. A Usage Policy bars unlicensed professional practice, military and weapons use, and operating critical infrastructure among other things. Downloads are gated behind a login and license acceptance; the card showed 199 likes and 402 downloads last month.

What others say

Meta’s claim is “most capable US open-weight model to date,” and that qualifier is doing a lot of work. Artificial Analysis’s independent Intelligence Index (v4.3.2) puts Muse Spark Open at 39, with the page noting it is “well above average among comparable models (median: 19).” The comparison table on the model card itself lists GLM 5.3 at 45, Kimi K3 at 44, Qwen3.8-2.4T-A95B at 40 and Mistral Large 4 at 38 — so on a global reading of the same table, “strongest” is fifth place. Engadget points out that the benchmarks where Meta reports wins over GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning are all Meta’s own, not independent results, and the card contains counterexamples: OSWorld Verified at 80.4 trails Kimi K3’s 84.8, while MCP-Atlas at 89.2 beats Qwen’s 84.5. LMArena added the model the same day but published only “scores coming soon,” so there is no Elo to quote. The counter-view on cost came from the same outlets: Engadget’s headline leads with “1TB of RAM,” its reporting puts four 256GB Mac Studios at roughly $38,000, and Apple still has not published a price or ship date for the 512GB configuration beyond “late October.” The piece concludes that this class of hardware is “pretty squarely out of reach for the typical vibe coder.” Apple’s framing is the engineering one: the same MLX post benchmarks Muse Spark Open’s decode speed against Qwen 3.8 and Mistral Medium 3.5 and calls it comparable to those dense models.

Our take

The real barrier here is not the download, it is memory and money. 1.64TB of weights, roughly 493GB after NVFP4, means the practical self-hosting entry point is two maxed-out Mac Studios, and Apple has not priced the 512GB configuration yet; the HF card’s “700GB of VRAM” line pushes the other route toward a multi-GPU server. Engadget’s $38,000 for four 256GB machines is already in small-cluster territory, which is why $1.25/$4.25 on an API is the more realistic option for an individual developer — and why Meta seeded seven platforms at once. The second judgment is to treat the vendor benchmarks as a claim, not a verdict: a 39 on an independent index, an OSWorld loss to Kimi K3 and an MCP-Atlas win over Qwen are all true at once, which points to strength in tool calling and agentic loops rather than across-the-board intelligence. Still unknown are LMArena’s blind scores, the price fields Artificial Analysis has not populated, and how Meta will count “affiliates” against that $20M threshold.

How to try it

The API route is shortest: search the model name on OpenRouter and pay $1.25/$4.25 per million tokens. Meta’s blog lists Baseten, Fireworks, Groq, Modal, Nebius and Together AI as well, and the PyTorch post says SGLang, vLLM, TensorRT-LLM, Transformers, Unsloth and llama.cpp already support it. For the local route, follow Apple’s setup — two M5 Ultra Mac Studios, the macOS 26.2 Thunderbolt 5 RDMA driver, MLX v0.30.1 or newer — then run the benchmark the post ships:

mlx.launch --hostfile hostfile.json --backend jaccl --env MLX_METAL_FAST_SYNCH=1 \
  -- mlx_lm.benchmark --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 -p 8192 -g 128 -b 1

For a local agent loop, Apple points at mlx_lm.server --port 8080 and http://127.0.0.1:8080/v1. To fine-tune, the PyTorch post points at TorchTitan, which ships a model definition for Muse Spark Open with multi-turn SFT, LoRA and TitanRL.