Muse Spark Open: What Meta's 816B Open-Weight Release Actually Gets You

Q: Can I run Muse Spark Open locally on a Mac Studio?

Yes with constraints: two 512GB M5 Ultra Mac Studios via Thunderbolt 5 and MLX (about $22,000 of hardware), 36.7 tokens per second on four machines or 25 on two at 8k context with NVFP4. Full 1M context needs about 266GB of extra KV-cache memory. Two-bit GGUF quantizations fit in about 300GB of combined RAM and VRAM.

Q: Is Muse Spark Open free for commercial use?

For most companies yes: the Muse Model License allows free commercial and internal use, but organizations exceeding $20 million aggregate gross revenue in any 12-month period must sign a separate written agreement with Meta to offer hosted model-as-a-service to third parties. Downloads are gated behind an approved Hugging Face request. It is open-weight, not open-source under OSI definitions.

Q: Should my business use Muse Spark Open or a hosted API?

Default to the hosted API at $1.25 per million input and $4.25 per million output tokens via OpenRouter, Fireworks, Baseten, Together AI and others. Self-host only if contracts mandate data residency, prohibit third-party egress, or require non-Chinese model provenance. If provenance is flexible, DeepSeek, GLM, Qwen, or Kimi open weights deliver higher capability per dollar on the same hardware.

Last Updated: October 16, 2026

Meta released the weights of Muse Spark Open on October 15, 2026. On paper it is a landmark: 816 billion parameters, 1 million token context window, runs on two Mac Studios on your desk. It is also, by Meta's own citation of an independent index, the weakest member of Meta's own model family. This post unpacks what actually shipped, what it costs to run, where it wins, and whether it belongs in your stack.

Key Takeaways

Muse Spark Open is an 816B total / 42B active parameter mixture-of-experts model with a 1M token context window, natively multimodal (text, image, video, audio input; text output), released October 15, 2026 under the custom Muse Model License.
Meta's "most capable US open-weight model" claim is geographically honest: on the independent Artificial Analysis Intelligence Index it scores 38.9, below GLM 5.3 (44.8), Kimi K3 (43.6), Qwen3.8 2.4T (39.9), and DeepSeek V4.1 Flash (39.5).
It is the weakest member of Meta's own family: the hosted Muse Spark 1.3 scores 61 on the same index. The open release is not the flagship.
The desktop story needs ~$22,000 of hardware (two 512GB M5 Ultra Mac Studios) plus ~266GB more for the KV cache at full 1M context; Apple's measured speed is 36.7 tokens per second on four machines.
Where it genuinely wins: tool orchestration and safety. JobBench 61.8 and MCP-Atlas 89.2 are row-best vs GLM/Qwen/Kimi, and its StrongREJECT adversarial score of 0.5 is roughly 125x better than the nearest peer.
Practical bottom line: hosted API first ($1.25/$4.25 per million tokens in/out), Chinese open-weights for price-sensitive local work, and Muse Spark Open only when compliance or data residency makes "not a Chinese lab" a procurement requirement.

What Shipped on October 15

Answer first: Muse Spark Open is the open-weights release of Meta's Muse Spark model. It is a sparse mixture-of-experts transformer: 816 billion total parameters with 42 billion active per token (top-8 routing across 256 experts), 62 layers, a 1,048,576-token context window, and finish, a 1.8B perception encoder that handles images, video, and audio. It takes multimodal input and produces text output. The weights come in BF16 (~1.6TB) and NVFP4 (~530GB) flavors from Meta, with quantizations from AMD, Intel, and Unsloth following within a day.

The release sits inside a crowded week. Meta shipped Muse Spark 1.2 with the Muse Code terminal agent on August 5 and 1.3 in September, both closed. Zuckerberg promised open weights for 1.2 in August "soon," then again for 1.3 in September. What arrived on October 15 is neither: it is a checkpoint derived from the August Muse Spark base with additional safety post-training, published under a new Muse Model License rather than an OSI-style open-source license. Meta's framing: "the most capable US open-weight model to date." Note the word US.

Inference providers listed at launch: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights are additionally available through Amazon SageMaker and the GCP Model Garden. Day-one inference framework support covers SGLang, vLLM, TensorRT-LLM, Unsloth, Hugging Face, and llama.cpp, plus MLX on Apple silicon. Meta shipped its own Helion GPU kernels into SGLang with claimed kernel-level speedups of 1.03x to 3.95x across NVIDIA and AMD parts.

A release-date coincidence worth noting: the Human aware AI news thread on Hacker News reached 855 points and 427 comments within about 19 hours, with a recurring complaint that the released checkpoint is not the promised flagship. More below.

The Family Paradox: Meta's Open Model Is Its Weakest Model

Answer first: On the independent Artificial Analysis Intelligence Index, Muse Spark Open scores 38.9. The hosted Muse Spark 1.3 scores 61. Every serious open-weights competitor outranks it: GLM 5.3 at 44.8, Kimi K3 at 43.6, Qwen3.8 2.4T at 39.9, DeepSeek V4.1 Flash at 39.5. "Most capable US open-weight model" is doing precise work in that sentence, and it is a claim about borders, not capability.

This is genuinely unusual in the open-weights era. Usually the open release IS the flagship: DeepSeek, Qwen, Kimi, and GLM open their best models or ship open and closed versions that are close in capability. Meta's October release inverts the pattern. The open checkpoint trails the hosted flagship by roughly 22 index points and falls behind every leading Chinese open-weights release on Artificial Analysis's snapshot from launch day.

Understand this omission in light of Meta's positioning: it puts Meta first in a category defined so the comparison set excludes the labs it trails. After 16 months without an open release following Llama 4, and after an explicit public promise in August that Spark 1.2 weights were coming "soon" and a September repeat for 1.3, shipping something else under a custom license is a decision worth naming as such. HN put it bluntly, with user halJordan writing that Meta "promised to open source muse 1.2 which is good. And they give us this bait and switch."¹

Worth keeping the counterbalance: Meta has form on honest benchmark tables. When it open-sourced Muse Glimmer (30B, Apache 2.0) on August 10, its model card printed all 26 comparison rows, including the 13 where it lost, and named the winner of each. Spark Open's card is more selective in what it leads with, but the methodology report is public.

The Desktop Illusion: What "Runs on Two Mac Studios" Actually Costs

Answer first: The two-Mac-Studio setup is real but constrained. It requires two M5 Ultra configs with 512GB each (the 512GB tier ships late October at $5,499 base per machine, roughly $22k for the pair assembled for this workload), Apple's measured 36.7 tokens per second on four machines (25 on two), and a KV cache that adds ~266GB at the full 1M context window. For comparison, a dual DGX Spark box runs DeepSeek V4 Flash around 60 tokens per second for a fraction of the price.

Start with the numbers themselves. The Apple MLX engineering post (published the same day as a partner) measured:

36.73 tokens/second decode on four M5 Ultra Mac Studios (NVFP4, batch 1, 8,192 token context), described by Apple as "well above average human reading speed"
24.99 tok/s on two machines (tensor parallel), 15.74 on two (pipeline parallel)
Time to first token: 4.8s on four machines at 8k context, scaling to 40-50s at 65,536 tokens on a two-machine rig
M5 Ultra vs M3 Ultra: up to 4.6x faster time-to-first-token, 1.6x faster decode, power draw 200W→300W. Apple's paper includes the magic footnote "a fraction of the power used by a common toaster," which is the most Meta-collab sentence in the whole post.²

The memory math explains why the desktop story has asterisks. The HF card's own arithmetic: 2 (K and V) x 16 KV heads x 64 head dimension x 62 layers = 126,976 values per token, about 250KB per token at BF16. At the full 1,048,576-token context window that is ~266GB, roughly half a Mac Studio's memory, before your prompt. The "1M context on your desk" row is therefore aspirational for most buyers: you can hold the NVFP4 weights in ~1TB across two machines, but a genuinely 1M-token working set needs a third machine or aggressive quantization plus context management. Practically, most desk deployments will run 128K to 256K context. (The DSpark speculative drafter and Meta's sliding-window layers amortize some of this; the model ships an adaptive checker that reads the model's sliding-window attention to keep drafter memory cheap at long context.)

Then there is the receipt: the M5 Ultra Mac Studio starts at $5,499, but the two-machine Muse Spark Open setup needs the 512GB tier (shipping late October), and community estimates put the pair around $22,000. Add the KV cache problem and the honest deployment envelope is: $22k buys you roughly 25 tok/s at mid-range context with a model that scores 38.9 on the AA index.

For context on what the same kind of money buys: a dual DGX Spark (GB10) desk rig runs DeepSeek V4 Flash and GLM 5.3-Flash-class models at 50 to 60 tok/s, and the HN operator colingauvin reports 50-60 tok/s runs of GLM 5.3-Flash with 1M context on a dual-Spark setup at ~4bpw EXL3. If your goal is "frontier-shaped model on my desk," the honest 2026 answer is that Chinese open-weights families on the same desk hardware outperform the Мета checkpoint on both speed and index score. The Spark Open desktop pitch is not really a price-performance pitch; it is a sovereign-provenance pitch.

When the desk rig is legitimately rational: Australian health, legal, and defence-adjacent contracts that demand data residency, no third-party egress, and per-query metering kept away from a US hyperscaler's logs all start to bend the math. If the compliance clause says the model cannot be "hosted by a foreign provider" and the workload is agentic tool-orchestration inside a private VPC, the $22k rig becomes a line item rather than a vanity purchase. That is a niche. It is a real niche.

Where It Genuinely Wins: Tooling and Safety

Answer first: Strip the index optics and the model is optimized for exactly one job profile: agentic tool-orchestration with a compliance-grade safety posture. On Meta's published table (xhigh reasoning, temperature 1.0), it leads all open-weight peers on JobBench (61.8) and MCP-Atlas (89.2), tops the no-tools knowledge probes, and posts the strictest adversarial-safety numbers of any model in the comparison, with a StrongREJECT v2 adaptive score of 0.5 versus 62.9 for the nearest competitor.

Read Meta's benchmark table row by row and a pattern emerges. The rows where Spark Open wins are orchestration and professional-delegation rows; the rows where it loses are raw reasoning and coding (Terminal-Bench 2.1: 76.4 vs Kimi K3's 85.0; DeepSWE 1.1: 60 vs Kimi's 68). The launch demos line up with the same profile: Meta showed Muse Spark Open shipping real iOS changes to Ice Cubes, the open-source Mastodon client, and end-to-end multimodal work in DaVinci Resolve (frame-by-frame review, auto vertical crops).

The safety numbers deserve their own beat. StrongREJECT v2 (adaptive) is lower-is-better, and Muse Spark Open posts 0.5 where the best peer in the table posts 62.9. On FORTRESS adversarial, attack refusals sit at 5.8, versus 28.2 for GLM 5.3 and 26.7 for Kimi K3. Meta's internal cyber benchmarks show 99.2% jailbreak refusal and MASK-benchmark honesty at 98.1%. Additional safety post-training is explicitly cited as the delta between the open checkpoint and its hosted base, which is the kind of unglamorous engineering that agentic fleets actually need when you hand a model a shell and a credit card.

One caveat with teeth: benchmark harnesses are not controlled experiments. Per Meta's methodology notes, DeepSWE rows compare per-model harnesses (Muse Code for Spark, Claude Code for Qwen, mini-swe-agent for GLM, Kimi Code for Kimi), so those are complete-system numbers, not apple-to-apple model measurements. The same report acknowledges judge selection is a measurement dependency.³

The License: Open Weights, But Read the Footnotes

Answer first: Muse Spark Open ships under the Muse Model License, not an OSI-approved open-source license. Downloads are gated behind a Hugging Face request form. If your company and affiliates exceed $20 million gross revenue in any 12-month period and you offer model-as-a-service, you must negotiate a separate written agreement with Meta. Internal use and research are exempt.

This is the same shape as the Kimi/Kuaishou-style licenses and the older Llama community licenses. For most consultancies below the revenue threshold, the practical summary is: free to run in-house, free to build on, cannot resell the model as an API if you are big, and re-downloads require an approved HF request (users report approval times near 10 minutes). HN user popinman322 flagged the market effect: large inference providers facing a separate license negotiation can reduce hosting price competition, a dynamic that played out with Kimi K3's hosting terms.

Is it open source? No, and HN argued about definitions at length: the weights are free to inspect and build on for most orgs, but "open weights" and "open source" are legally different things. If you work with clients who care about that distinction (procurement increasingly does), call it "open weights" in writing.

Should You Run It Locally? A Decision Framework

Answer first: For nearly everyone: no, not as the default. Read the hosted API first at $1.25 per million input and $4.25 per million output tokens. Run weights locally only if data residency, no-egress commitments, or "not a Chinese lab" procurement language forces it. Otherwise, the same desk hardware runs GLM 5.3-Flash or DeepSeek V4 Flash faster and cheaper.

The practical decision tree:

Compliance requires local + provenance (health, legal, government-adjacent, client contract says no foreign hosting): Muse Spark Open on two 512GB Mac Studios or a GPU box is a defensible choice, with the caveat that 1M context costs ~266GB of extra memory at BF16. Budget real engineering time for memory budgeting and guardrails; Meta's card explicitly recommends deploying with additional guardrails.
Compliance requires local, provenance is flexible: GLM 5.3-Flash or DeepSeek V4 Flash on a dual DGX Spark or Mac Studio rig beats Spark Open on price-performance today (50-60 tok/s vs 20-40 tok/s, and those model families open their flags with Apache/MIT-class licenses).
Cloud is fine: Do not run any of this locally. Hosted API access costs $1.25/$4.25 per million tokens via OpenRouter, Fireworks, Baseten, Together, Modal, Nebius, or Groq. For price-sensitive scale, DeepSeek and GLM APIs undercut Meta's; for agentic criticality, the interesting question is whether Spark Open's tool-safety profile justifies the premium in your deployment.

A note on pricing psychology: at launch the hosted Spark Open costs the same as the hosted Spark 1.3 (identical $1.25/$4.25 list pricing). If the cloud is an option, the rational default is to rent the stronger checkpoint. The open weights' value proposition is control, not quality; that's the entire reason to reach for them.

Australian Angle: Where This Actually Fits a Flowtivity Client Stack

Answer first: For Australian SMBs, Muse Spark Open is a specialist tool. Its realistic role is "the compliance-approved open-weights model" for sectors where the US provedance of weights and the tightest refusal behavior make procurement comfortable: health, legal, fintech, government-adjacent. For general agent workloads, hosted or Chinese-open alternatives deliver more capability per dollar.

We run DeepSeek V4 Flash on dual DGX Spark hardware at Flowtivity, so the question we get is "should we run Meta's 816B instead?" The honest answer from our bench: not for capability (Spark Open scores below DeepSeek V4.1 Flash on the AA index), not for speed (25 tok/s on two $11k Macs vs 60 tok/s on one desk rig), and not for price ($1.25/$4.25 hosted vs DeepSeek API's commodity rates). The one yes is procurement: when a client's contract language effectively excludes Chinese-origin weights and the workload is agentic, Spark Open's combination of row-best tool benchmarks and the strictest safety scores of its peer set is a legible story to a risk committee that a $300 GPU box running an Apache-2.0 Chinese model is not.

Watch for a smaller sibling. The Community wants a ~300-400B total / 15-20B active Spark variant (an HN wishlist item" during launch week); a Glimmer-class distillation of 1.3 would change the local calculus immediately. If Meta ships it with the same day-one tooling, the desk story gets interesting for anyone who can't justify $22k.

Frequently Asked Questions

What is Muse Spark Open?

Muse Spark Open is Meta's open-weights release of its Muse Spark model, published October 15, 2026. It is a mixture-of-experts LLM with 816 billion total parameters (42 billion active per token), a 1 million token context window, and multimodal input (text, image, video, audio) with text output. It runs via APIs (Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together AI) or locally via MLX on Apple silicon, SGLang, vLLM, TensorRT-LLM, llama.cpp, or Unsloth.

How good is Muse Spark Open compared to other open-weight models?

On the independent Artificial Analysis Intelligence Index it scores 38.9, behind GLM 5.3 (44.8), Kimi K3 (43.6), Qwen3.8 2.4T (39.9), and DeepSeek V4.1 Flash (39.5). Its strongest rows vs peers are agentic tool use (JobBench 61.8, MCP-Atlas 89.2, both best-in-table) and safety (StrongREJECT v2 0.5 vs 62.9 next best). Meta's own hosted flagship Muse Spark 1.3 scores 61 on the same index, so the open release is the weakest member of Meta's current family.

Can I run Muse Spark Open locally on a Mac Studio?

Yes, with constraints. Apple and Meta demonstrated it on two 512GB M5 Ultra Mac Studios linked via Thunderbolt 5 with MLX (about $22,000 of hardware); Apple measured 36.7 tokens per second on four machines and 25 on two at 8k context with NVFP4 quantization. The full 1M context window requires roughly 266GB of extra KV-cache memory beyond the weights, so most desk deployments run shorter contexts. Two-bit GGUF quantizations fit in about 300GB of combined RAM and VRAM. Cheaper rigs (a single 512GB Mac, dual DGX Spark) run the model quantized but slowly.

Is Muse Spark Open free for commercial use?

For most companies, yes. The Muse Model License allows free commercial use and internal deployment, but companies whose aggregate gross revenue (including affiliates) exceeds $20 million in any 12-month period must sign a separate written agreement with Meta to offer the model as a hosted service to third parties. Downloads require an approved Hugging Face access request. It is open-weight, not open-source under OSI definitions.

Should my business use Muse Spark Open or a hosted API?

Default to the hosted API ($1.25 per million input tokens, $4.25 per million output via OpenRouter, Fireworks, Baseten, Together AI, and others). Run the weights yourself only if your contracts mandate data residency, prohibit third-party egress, or require non-Chinese model provenance. If provenance is flexible, Chinese open-weights families (DeepSeek, GLM, Qwen, Kimi) deliver higher capability per dollar on the same local hardware.

How and where did AJ Awan see this landscape?

AJ Awan is the founder of Flowtivity, a Gold Coast AI consultancy, and a former EY enterprise architect (TOGAF). Flowtivity runs open-weights models on local dual-DGX-Spark hardware daily, so this analysis is written from implementation experience, not launch-week enthusiasm.

Sources

Meta AI research blog, "Muse Spark Open" (research.meta.ai/blog/muse-spark-open), Oct 15 2026; and Meta developer blog, "Build with Muse Spark Open" (dev.meta.ai/resources/blog/build-with-muse-spark-open), Oct 15 2026.
Hacker News discussion of Muse Spark Open (news.ycombinator.com/item?id=50107379): user halJordan on the 1.2 promise; user colingauvin on dual-Spark GLM 5.3-Flash at 50-60 tok/s; user popinman322 on hosting-license effects; confidence flags per ResearchAudio's GothamLights methodology. (Thread: 855 points, 427 comments at ~19h.)
Meta model page benchmark table and evaluation methodology report (dev.meta.ai/models/muse-spark-open), xhigh settings, temperature 1.0; Apple Machine Learning Research post "Distributed Inference with MLX" (machinelearning.apple.com/research/mlx-distributed-inference), Oct 15 2026. Apple and Meta co-marketed the launch; Zuck/Ternus X exchange via The Verge (Jay Peters, Oct 15 2026).
Hugging Face model card meta-models/Muse-Spark-Open-816B-A42B (architecture, KV cache arithmetic, license, gating, release date Oct 15 2026); Artificial Analysis v4.3.2 snapshot Oct 15 2026 (index 38.9, peer scores, Terminal-Bench 4.0 8.1) via kingy.ai launch-day deep dive (Curtis Pyke) and artificialanalysis.ai. Terminal-Bench 2.1 vs 4.0 are different task sets: do not compare directly.
License terms: Muse Model License via Hugging Face and secondary coverage (igeeksblog, Oct 15-16 2026, $20M 12-month MaaS clause); pricing via OpenRouter model page (openrouter.ai/meta/muse-spark-open).
Local quantization sizes: Unsloth "Meta Muse Spark Open: Run Locally" guide (unsloth.ai/docs/models/muse-spark-open, 2-bit GGUF ~300GB) and Hugging Face engineering post (3-bit oQ3e ~344GB on 512GB M3 Ultra, 21 tok/s generation); BF16 ~1.6TB and NVFP4 ~530GB per Meta model card.

Disclosure: Flowtivity runs DeepSeek V4 Flash on dual DGX Spark desks; no affiliation with Meta or Apple. Benchmarks cited are launch-window numbers and will move.