Meta Muse Spark Open Review (2026): The US Open-Weights Answer to Kimi K3

Meta published 816 billion parameters of frontier-adjacent weights and optimised them for two Mac Studios. That is a real first. It is also not the best model you can download this week.

By AI Tool Scout · Published October 16, 2026 · 22 min read

What Is Muse Spark Open?

Muse Spark Open is the open-weight release of Meta's Muse Spark model, published by Meta Superintelligence Labs on October 15, 2026. The launch post is blunt about what it is: "the most capable US open-weight model to date", a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window, "built for coding and agentic work" and "long-horizon tasks, complex software engineering, and multi-step reasoning." The weights are on the meta-models Hugging Face organisation as Muse-Spark-Open-816B-A42B, in BF16 and NVFP4, and the model card credits Meta Superintelligence Labs as authors with an October 2026 release date.

The architecture is a sparse mixture-of-experts transformer, and the model card is unusually specific about it. 62 layers, hidden dimension 8,192, 256 experts with 8 selected per token by top-k routing, 128 query heads and 16 key-value heads (grouped-query attention with a head dimension of 64, QK-norm and an attention output gate), a vocabulary of 202,048 tokens with untied input and output embeddings of about 1.66B parameters each, and a 1.8-billion-parameter perception encoder with 50 layers handling vision. Input is text, images, video and audio; output is text only, which matters — this is not a generator of images, video or speech, and a product that needs those has to bolt them on. The context window is set at max_position_embeddings = 1,048,576, shared across text and the media tokens. Meta recommends temperature 1.0, top_p 1.0, top_k disabled and a reasoning effort of high or xhigh, with xhigh as the default.

The lineage matters for anyone trying to read a press release accurately. Meta shipped Muse Spark in April 2026 — its first model release since Llama 4, and a clean break from the Llama line, which had fallen behind. That first Spark was hosted only. On August 10, 2026 Meta released Muse Glimmer, a 30B model under a fully permissive Apache 2.0 licence, and Zuckerberg said the weights for Muse Spark 1.2 would follow "soon". Spark 1.3 then shipped on September 2, still closed, with another public promise that open weights were coming. On October 12, Meta announced the closed Muse Wonder family. Three days later it finally published the weights — and what it published is Muse Spark Open, a model whose base is shared with the hosted Spark 1.3, not a re-release of 1.2. That is roughly a six-week gap between promise and delivery, and plenty of developers noticed; the delay is the single most-repeated complaint in the launch threads.

The Apple Angle: Two Mac Studios, M5 Ultra, and John Ternus

This is the part of the launch that got the headlines, and it is genuinely unusual: Meta and Apple shipped a joint optimisation. Meta's post says 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", and the framing is directly Apple's: Apple's unified memory architecture lets the M5 Ultra GPU address up to 512 GB per Mac Studio, so two machines can hold the whole model, and the chip's 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. Meta's summary line is the one everyone quoted: "two Mac Studios behave like one supercomputer with a terabyte of unified memory."

The political subplot is real. Zuckerberg posted on X: "Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple." John Ternus became Apple's CEO in September 2026, taking over from Tim Cook, and he replied publicly. Two companies that spent a decade fighting over privacy advertising and app-store rules co-signed a frontier model launch, and Meta's AI chief Alexandr Wang told CNBC the reason in plain terms: many capable open-weight models are simply too big for most personal computers, and the Mac Studio's memory is what makes a model this size practical to run at all.

It is not only marketing, though it is partly marketing. Apple's Machine Learning Research team published the same day a detailed post on local and distributed inference with MLX and RDMA over Thunderbolt 5, and the numbers are real and checkable. Apple measures three parallelisation schemes on M5 Ultra at an 8,192-token context: 15.74 tokens per second on two Mac Studios using pipeline parallelisation, 24.99 on two using tensor parallelisation, and over 36 tokens per second on four — which Apple notes is "well above average human reading speed." It also publishes time-to-first-token and power figures: the M5 Ultra's GPU Neural Accelerators make prefill up to 4.6× faster than an M3 Ultra, decode is 1.6× faster because bandwidth rose from 819 GB/s to 1.2 TB/s, and peak power is about 300 W for the M5 Ultra against about 200 W for the M3 Ultra — a fraction of a GPU node, and the reason Apple can argue a Mac Studio belongs in any office. The connectivity is the unglamorous engineering that makes it work: macOS 26.2 added a low-latency RDMA driver over Thunderbolt 5, MLX 0.30.1 and later builds collectives on top of it, and Apple's demo machine used two Thunderbolt cables to double link bandwidth.

Meta's on-device demos are the appetising part and worth taking at face value with one caveat. The company says it tested agentic workflows running entirely locally with Muse Code powered by Muse Spark Open — shipping real iOS changes (bug fixes, new features, a full build) to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end while reviewing footage frame by frame to find the strongest moments and crop them vertically. That is a genuine demonstration of long-context agentic work on a desktop. The caveat is the numbers underneath it: 15 to 25 tokens per second on the two-machine setup Apple benchmarked is roughly a sixth to a quarter of the ~102 tokens per second the hosted API delivers. A local agent that reads a large codebase and reasons for hours at that speed is a different experience from a hosted one, and no demo video conveys it.

Key Features: Coding, Agentic Work and Long-Horizon Tasks

Four things define this release, and each one has a limit attached.

• Tool use is the genuine win. MCP-Atlas 89.2 beats every model in Meta's comparison table, the Chinese ones included: Kimi K3 at 82.3, GLM 5.3 at 84.2, Qwen3.8 at 84.5. The benchmark gives a model plain-language tasks without naming the tools, and the model has to work out which of 220 tools across 36 live MCP servers to call and in what order. That is the closest public test to how agent frameworks actually fail — not by being unable to write code, but by picking the wrong tool. Meta's Toolathlon Verified 72.8 is competitive (Kimi K3 76.5, GLM 5.3 73.0) rather than best, so read MCP-Atlas as a strong signal, not a sweep.
• Long-horizon agentic work, on Meta's own evidence. JobBench 61.8 — professional tasks workers would delegate, including reconciling contradictory records — is roughly double Thinking Machines' Inkling (31.1) and Nvidia's Nemotron 3 Ultra (27.2), and edges GLM 5.3's 61.4 and Kimi K3's 54.3. That is the strongest single number in the launch for the "runs for hours and finishes" claim. It is also the number to hold loosely: JobBench comes from the official leaderboard, whose current configuration specifies OpenCode v1.14.18 as the agent and a Grok 4.3 rubric judge, and Meta's methodology document says its runs on third-party models were not tuned for each of them. A 61.8-to-61.4 comparison is not a decisive advantage.
• A million-token context over four modalities. Text, image, video and audio share the full 1,048,576-token window, so a single prompt can hold a codebase, a screenshot, a screen recording and a voice note. The multimodal rows are mixed rather than dominant: HLE with tools 61.0 against Kimi K3's 56.0, IntPhys2 68.3 against 60.3, SimpleVQA with tools 73.6 against 71.9, but OSWorld Verified 80.4 against Kimi K3's 84.8, BabyVision with tools 80.2 against 85.7 and VideoMME v2 61.5 against 65.4. Note also that VideoMME v2 in this evaluation uses subtitles without audio, so it does not measure audio comprehension, and ZeroBench's 50.0 is pass@5 — success across up to five samples — not single-answer accuracy.
• Safety post-training, and the strongest jailbreak numbers in its own table. Meta is explicit that the open version "has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails." On StrongREJECT v2 (adaptive), where lower is better, Muse Spark Open scores 0.5 against Kimi K3 at 66.6, GLM 5.3 at 62.9 and Mistral Large 4 at 65.6. On FORTRESS adversarial (static) it records ARS 5.8 / ORS 17.4, the best in the table. On cyber misuse with adaptive attackers and no system safeguards, Meta says it refused 99.2% of jailbreaks on an internal benchmark and 94.8% across PolyGuard, RedCode and CyberBench 2024, where Kimi K3 refused 75.5% and 70.2%. Biology and chemistry refusals are above 97–98% across single-turn and multi-turn agentic benchmarks. Meta says the model met the thresholds of its Superintelligence Scaling Framework for release, with the full evaluation report "available soon" — a phrase worth noticing, because the detailed report was not published at launch. Meta's own advice is also a limitation: it "strongly recommends deploying Muse Spark Open as part of an AI system with guardrails tailored to your intended use cases", which is a vendor telling you the unguarded weights are not the finished product.

Benchmarks vs Kimi K3, GLM 5.3 and Qwen3.8

There are two tables in circulation and they do not tell the same story. Meta's is built from its own runs plus credited third-party evaluators; the independent one is Artificial Analysis. Here is both, with the caveats attached.

What Meta's own table says

Muse Spark Open is measured at xhigh reasoning effort. Against the Chinese open models — the comparison that actually matters, since that is the market this release is aimed at — it wins the agentic and knowledge rows and loses the coding rows:

Four disclosures change how you read that. First, Meta's table is an earlier snapshot: it takes Artificial Analysis values dated October 7, and the index has moved since. Second, the harnesses differ by design — Meta identifies Muse Code for Spark Open, Claude Code for Qwen, mini-swe-agent for GLM and Kimi Code for Kimi on DeepSWE, so each column describes a complete evaluated system rather than a controlled comparison. Third, Meta credits Scale for MCP-Atlas and HLE-Diamond and the official JobBench leaderboard for JobBench, and the public evaluator pages had not yet listed Muse Spark Open when these figures were published — so they are Meta-published results with attribution, not verified public leaderboard positions. Fourth, the ratings and percentages should never be averaged together: GDPval is an Elo rating and AA Omniscience uses its own score, which rewards correct answers, penalises wrong ones and leaves abstention unpenalised, on a scale from −100 to +100. Muse Spark Open's 25 there is a knowledge-reliability score, not 25% accuracy and not a hallucination rate; its separately reported accuracy is 39.3%.

The independent index — and the coding number that hurts

On Artificial Analysis Intelligence Index v4.3.2, the composite that now weights ten evaluations toward agentic and knowledge work, Muse Spark Open at xhigh scores 39, placing it #10 of 119 models in its class and well above the comparable-model median of 19. It is also #19 of 77 on speed at 102 output tokens per second — notably fast — and fairly concise at 110M tokens to run the index against a median of 140M. The weak spot is cost: it ranks #32 of 41 on price, and Artificial Analysis's own summary says the model is "amongst the leading models in intelligence, but particularly expensive when comparing to other open weight models of similar size." The context for that judgement is stark: $1.25 input against a median of $0.43, and $4.25 output against a median of $1.59. Its measured cost per index task is $0.84. Cheaper open routes exist by a wide margin, including DeepSeek V4.1-Flash at roughly $0.15 and $0.60 off-peak and GLM 5.3 at $1.40 and $4.40, both covered in our LLM API cost guide.

The single most damaging independent number is inside the current index itself. Terminal-Bench 4.0, which the live composite uses, puts Muse Spark Open at 8.1% against 41.9% for GLM 5.3, 59.6% for Claude Opus 5.5, 59.1% for GPT-6 Astra and 63.6% for Claude Sonnet 5.5. This does not contradict the model's headline 76.4 on Terminal-Bench 2.1; the two versions use different task sets — 66 tasks for 4.0 against 89 for 2.1 — and neither their difference nor their ratio measures a change in the model's ability. But it does mean that on the benchmark the main independent evaluator currently counts, the coding-agent story is far weaker than the launch press suggested. The rest of the independent profile is mixed in the same way: HLE without tools 43.9% against 61.4% for Opus 5.5, 57.1% for Gemini 4 Argon and 52.9% for Sol 6.1, while SciCode 58.4% exceeds Astra's 56.5% and Sol's 54.2% and trails Opus's 66.9%. A model can be genuinely useful in scientific coding without matching the frontier across the suite, and that is the accurate description here.

What "most capable US open-weight model" actually claims

The phrase is doing real work and it is narrower than it sounds. Against US open models, Meta's lead is large and uncontested: Inkling at 26 on the index, Nemotron 3 Ultra at 23, and Muse Spark Open at 39, with Meta often ahead by 20 points or more on individual rows. Against the world's open models it is not first: GLM 5.3 (45), Kimi K3 (44), MiMo-V2.6-Pro (46) and Qwen3.8 (40) all score higher on the current index, and it trails Kimi K3 and GLM 5.3 specifically on the coding and terminal rows. The geographic qualifier also matters for procurement without settling anything about quality — and where a model is developed does not determine where your data is processed, since OpenRouter already offers US hosting for Chinese models including DeepSeek V4 Pro and Kimi K3. Meta's own comparison table makes the last point for us: its closed Muse Wonder Full scores 56 on the index against 39 for the model it gave away, so Meta kept roughly seventeen index points behind the paywall.

What It Costs to Run

Every figure below is approximate, in US dollars, checked on October 16, 2026 against Meta's launch post, model card and deployment documentation, Apple's MLX research post and Mac Studio specifications, the Hugging Face repository, the Muse Model License, Artificial Analysis, and launch coverage from Reuters, CNBC, The Guardian, The New Stack, VentureBeat and SiliconANGLE. Confirm current pricing, licence terms and hardware availability on the vendors' own pages before you build or buy.

The hardware bill: 4-bit still means half a terabyte

Start with the floors Meta publishes, because they are higher than the "runs on your desktop" headline implies. The model card states a requirement of a GPU with at least 700 GB of VRAM for the NVFP4 build and 2 TB for BF16. The deployment documentation gives weight footprints of about 537 GB for NVFP4 and about 1.65 TB for BF16, and Apple measured the 4-bit footprint at about 493 GB; none of those numbers includes the key-value cache, activations or runtime overhead, and NVFP4 is only about 30% of the BF16 size — quantisation made this model merely enormous instead of impossible.

Meta's own serving recipes are the authority here: NVFP4 runs on TP4/EP4 across four B300s and BF16 runs on TP8/EP8 across eight B300s. In Transformers, Meta says the BF16 checkpoint runs across four nodes of eight H200 GPUs with tensor parallelism 32 and the NVFP4 checkpoint runs on a single B200 node. Nothing in the consumer or prosumer category is close, and the desktop path is not a workaround for that — it is a second, entirely different machine class.

The break-even maths, which is the only number that decides this. At a blended API rate of about $0.78 per million tokens (a 7:2:1 cache-hit, fresh-input and output mix, derived from the published cache discount), the comparisons look like this: two Mac Studios at a rough $30,000 need about 38.5 billion tokens in total before their purchase price is repaid; a rented four-B300 node at an estimated $35,000 a month needs roughly 45 billion tokens a month to beat the API, or about 90 billion a month at the published eight-B300 rate of $70,000. A heavy individual user running 20 million tokens a day through the API spends about $15.60 a day. Most teams are more than an order of magnitude below the crossover, and the API also comes with no cluster, no operations burden, no expert-parallel debugging and no cold-start problem. The realistic middle path is a hosted endpoint you control the client for — Fireworks, Together AI, Baseten or Nebius — or, if data residency forces your hand, two Mac Studios under a desk, accepting roughly a sixth of the tokens per second.

One further cost that never appears in a table: running a 62-layer MoE with 256 experts across multiple machines is an infrastructure project. Expert-parallel sharding, all-to-all traffic, KV cache management at a million tokens and failure domains that cross hosts are real engineering. The Mac path hides most of that behind MLX and RDMA, which is exactly why Apple's involvement matters — and exactly why the two-machine demo is more interesting than a benchmark row.

Pricing and Availability

Everything below was live on October 15, 2026. Rates are approximate and move; check each vendor before you commit.

• APIs (day one): Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI — about $1.25 in / $0.15 cached / $4.25 out per million tokens, with the full 1M context on Together AI and Fireworks, and a 131,072-token ceiling plus organisation approval on Groq.
• Weights: free download from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, in BF16 (meta-models/Muse-Spark-Open-816B-A42B) and NVFP4 repositories, each shipping the tokenizer, chat template and processor config. No API key is needed once the weights land.
• Inference engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for NVIDIA, AMD, Apple and Intel silicon. The open-weight release ships custom, auto-tunable Helion GPU kernels in SGLang that Meta says deliver 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. For further tuning, Meta points to PyTorch's TorchTitan for SFT, LoRA and reinforcement learning.
• Desktop: Apple MLX, which pulls the weights directly from Hugging Face, on two or more M5 Ultra Mac Studios. The 512 GB configuration ships in late October 2026; the 256 GB configuration costs $10,799 and the entry 96 GB M5 Ultra starts at $5,499 and cannot reach 512 GB.
• The competition: Qwen3.8 is cheaper on every axis, Kimi K3 scores higher but costs more (~$3 / ~$15), GLM 5.3 lists ~$1.40 / ~$4.40 and scores higher, and DeepSeek V4.1-Flash is roughly a tenth of the output price.

Pros & Cons

5 Best Muse Spark Open Alternatives

Comparison Table

All prices are approximate, in US dollars, checked October 16, 2026, and change often. Confirm on each vendor's own pricing page and read each licence before you ship. Where a cached-input rate is shown as a dash, the vendor does not publish a headline cache discount on that model.

Which One Should You Choose?

• Use the hosted API if you want Muse Spark Open, and do not plan to self-host. At roughly $1.25 / $0.15 / $4.25 per million with an 88% cache discount, seven hosts and a flat 1M context, it is a fairly priced frontier-adjacent model — and about 102 tokens per second is genuinely fast. Push cache hit rates up first; that single lever is most of your input bill.
• Check Meta's Contributor tier before you build. Muse Spark 1.2 and 1.3 Contributor cost $0.10 and $0.20 per million against the open model's $1.25 and $4.25. If your data can be used for training, the hosted closed model is an order of magnitude cheaper than the open one, which is the strangest pricing fact of this launch.
• Choose Kimi K3 or GLM 5.3 if coding is your workload. Both score higher on the current index and both beat Muse Spark Open clearly on Terminal-Bench and DeepSWE, on Meta's own table. Meta wins on tool discovery and long-horizon agentic tasks; if that is your bottleneck, it wins for you, and if you are shipping code in a terminal environment, it does not.
• Only self-host above roughly 45 billion tokens a month. Two Mac Studios need about 38.5 billion tokens in total to repay a rough $30,000, and a rented four-B300 node needs about 45 billion a month to beat the API. Below that, the API wins on cost, latency, redundancy and the fact that someone else owns the pager.
• If privacy is the requirement, wait for the 512 GB Mac Studio to actually ship. The configuration Meta recommends is not on sale until late October and Apple has not published its price. Two 256 GB machines leave roughly 19 GB of headroom over the 493 GB NVFP4 footprint, which is not a comfortable place to run a 1M-token context.
• Do not buy this as a consumer product. There is no free chat tier, no Muse app integration and no consumer plan. If a personal assistant that remembers across weeks is what you want, that is the Muse Wonder tier, and if you want a small model that runs on the machine you already own, the honest answer is Qwen, Muse Glimmer or MiMo-V2.6-Pro.
• Read the licence before your legal team does. "Built with Muse Spark Open" in your UI and a $20M group-revenue trigger for anything that looks like model-as-a-service are the two clauses that decide whether this is free for you. The Kimi K3 comparison is the useful precedent, because the thresholds are nearly identical.
• Do not wait for the next model to decide. Meta promised these weights in August, shipped them in October, and simultaneously kept a 17-index-point stronger model closed. Build on what exists and treat the next open release as upside rather than a plan.

Frequently Asked Questions