Meta has opened the Spark. On 15 October 2026 it published the weights of Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42B active parameters per token and a 1-million-token context, calling it "the most capable US open-weight model to date". It shares one base model with Muse Spark 1.3 – the closed API model that powers the Muse agent – and weights are free to download, fine-tune and run, with a $20M revenue gate aimed only at companies reselling access.
Every other site will stop at "Meta goes open-source". The truth is more interesting: it is not open source (it's an open-weights custom licence), the raw capability numbers trail China's open flagships, and nobody's desk can hold it on one machine – even Apple's flagship model. What it does dominate is agent-safety testing, which may matter more for who actually ships it. Here is the honest picture.
Source note: specs, licence terms and benchmark numbers below come from Meta's launch post (15 Oct 2026), the model card and licence text in the official Hugging Face repo, Apple's Machine Learning Research measurements, Artificial Analysis and OpenRouter's public API listing. Benchmarks are Meta-reported (with Artificial Analysis's independent index where noted), not a VelsTech Solutions test – nobody outside Meta has independently measured an 816B MoE of this vintage yet beyond benchmark farms.
Related: the closed frontier line is covered in our Muse Wonder explainer, and the agent these models power is in our Muse personal-agent guide.
The short version: Spark Open in 60 seconds
- What: the open-weights twin of Muse Spark 1.3 – a sparse MoE reasoning model for coding and agentic work, natively multimodal (text, image, audio and video input; text out).
- Date / price: weights free to download worldwide from 15 October 2026 (HF, Google Cloud Model Garden, AWS SageMaker, Microsoft Foundry); hosted inference via 7 API platforms at $1.25/$4.25 per million tokens.
- Specs: 816B total / 42B active parameters, 1M-token context, 256 experts with 8 active per token; ships in BF16 (~1.6TB) and NVFP4 (~537GB) checkpoints.
- Headline claim: "the most capable US open-weight model to date" – a Meta claim scoped to US origin, not a global #1: GLM 5.3, Kimi K3 and Qwen3.8 all score higher on Artificial Analysis's independent index.
- Catch: two of them. The licence only bites on $20M+ model-as-a-service resellers, but the hardware gate is real: about 493GB at 4-bit means two 512GB Mac Studios or a B200 node – a single flagship Mac cannot run it.
- Who it's for: enterprises and builders who need a downloadable, safety-audited, US-origin agentic model and can fund the cluster – plus anyone fine-tuning off Spark's base. Consumer local-AI users stay on Muse Glimmer 30B.
What Muse Spark Open actually is
Three rungs now make up the Muse model family. Spark 1.1 → 1.3 are closed API models. Spark Open – the new rung – is built on the same base model as Spark 1.3, then given extra safety post-training so it behaves sensibly as a standalone model without Meta's system-level guardrails around it. And Muse Glimmer 30B, the Apache-2.0 model that runs on consumer hardware, was distilled from this same Spark lineage. In other words: the closed flagship, its open twin, and the developer-friendly small model.
Why now? Because the open-weights race went global. China's labs ship ever-larger open models (Kimi K3 at 2.8T total parameters is the largest open-weights model in existence), and Meta needed an answer that was both credible and different. Their differentiators: the safety work (more below), the 1M context on an agentic native multimodal base, and a licence designed as a middle ground between Apache-2.0 giveaways and old Llama-style user caps.
Specs at a glance
| Spec | Muse Spark Open | Muse Spark 1.3 (closed) |
|---|---|---|
| Parameters | 816B total / 42B active (sparse MoE, 256 experts, 8 active per token) | Undisclosed (used for the same family base) |
| Context window | 1M tokens | 1M tokens |
| Input modalities | Text, image, audio, video | Text, image, audio, video |
| Checkpoints | BF16 (~1.6TB) and NVFP4 (~537GB); AMD MXFP4/FP8 variants | API only – no weights |
| Independence | Runs without Meta's system guardrails (extra safety post-training) | Served inside Meta's Model API stack |
| Hosted price | $1.25 in / $4.25 out per million tokens (OpenRouter and peers) | $1.25 / $4.25 ($0.10/$0.20 Contributor mode) |
| Weights licence | Muse Model License (custom, free below $20M MaaS revenue) | Commercial API terms only |
The honest row above is the last-but-two: on Apple's own measurements the 4-bit checkpoint still occupies about 493GB of RAM at runtime (the NVFP4 file on disk is about 537GB, BF16 is over 1.6TB). That rules out not only every consumer GPU but also a single 512GB Mac Studio – the model was built to run across 2–4 clustered Mac Studios or a server node.
Benchmarks where it stands – and where it trails
Meta's model card compares Spark Open directly against the current open-weights field (all at xhigh reasoning effort, an apples-to-apples setting). Selected rows:
| Benchmark | Spark Open | GLM 5.3 | Kimi K3 | Qwen3.8-2.4T | Mistral Large 4 |
|---|---|---|---|---|---|
| AA Intelligence Index (higher better) | 39 | 45 ● | 44 ● | 40 ● | 38 |
| JobBench (agentic job tasks) | 61.8 ● | 61.4 | 54.3 | 52.7 | – |
| MCP-Atlas (tool calling) | 89.2 ● | 84.2 | 82.3 | 84.5 | – |
| HLE-Diamond, no tools | 26.1 ● | 16.4 | 22.2 | 17.3 | – |
| AA-Omniscience (knowledge/hallucination) | 25 ● | 14 | 20 | 4 | -5 |
| OSWorld Verified (computer use) | 80.4 | – | 84.8 ● | – | – |
| DeepSWE 1.1 (software engineering) | 60 | 67 | 68 ● | 57 | 62 |
| Terminal-Bench 2.1 | 76.4 | 83.9 | 85.0 ● | 82.0 | – |
| GPQA Diamond (science) | 89.0 | 91.7 | 93.5 | 93.5 ● | – |
| FORTRESS adversarial risk (lower better) | 5.8 ● | 28.2 | 26.7 | 21.6 | – |
(● marks the row's top score; "–" = not reported. Full table in the model card – Mistral Large 4, Inkling and Nemotron 3 Ultra round out Meta's comparison set.)
Read honestly, this shows a very specific profile. On raw knowledge and coding benchmarks Spark Open sits behind GLM 5.3 and Kimi K3 – DeepSWE 60 vs 68, Terminal-Bench 76.4 vs 85.0. It wins the agentic cluster (JobBench, MCP-Atlas, tool-calling) and the hallucination-fighting awards (AA-Omniscience 25 vs 4–20). And then there is the safety table, where it is not competitive – it is in a different league. On StrongREJECT v2, an adaptive jailbreak benchmark, Spark Open scores 0.5 while every rival scores between 62.9 and 86.4 (lower is better). Its Artificial Analysis index score of 39 ranks it around #10 overall – solidly frontier-adjacent, but the point is clear.
The most uncomfortable comparison for Meta, though, is internal. On the same Artificial Analysis index v4.3.2 where Spark Open scores 39, the closed Spark 1.3 scored 48 (as measured around the Muse Wonder launch). That is a plausible cost of the extra safety post-training and possibly of third-party serving configurations – Meta doesn't break down why, but if you're choosing between the twins, the closed API model is measurably smarter while the open one is measurably safer solo.
The licence: free until you resell
The weights carry the Muse Model License, and its core grant is generous: use, copy, modify, fine-tune, distil, merge, publish, distribute, sublicense and even sell – free of charge, for individuals and enterprises alike. It applies, notably, to models improved using Spark Open or its outputs too. It is not an OSI-approved open-source licence, though: two strings are attached.
- Attribution: show a prominent "Built with Muse Spark Open" notice in your product interface, documentation, site or settings; retain the copyright, licence and Usage Policy in copies and derivatives.
- The $20M gate: only if you run a Model-as-a-Service business – giving third parties API access to the model's outputs – and your aggregate gross revenue (company + affiliates, undefined term) tops $20M in any 12 months, you must then notify Meta and sign a separate agreement. Internal deployments, research and building the model into your own product are explicitly exempt.
| Condition | Muse Model License | Llama 4 Community | Apache-2.0 (Glimmer 30B) |
|---|---|---|---|
| User cap | None | 700M MAU | None |
| Revenue gate | $20M/12mo, MaaS only | None | None |
| Branding requirement | "Built with Muse Spark Open" | "Built with Llama" + Llama-prefix naming | Notices only |
| Derivative naming rule | None | Must start with "Llama" | None |
| EU restriction | None | Multimodal models excluded for EU companies | None |
In effect Meta targeted cloud resellers – the businesses that compete with its own Model API – while leaving product builders (SaaS features, internal agents, support bots) untouched at any size. Two practical notes: "affiliates" is undefined, so conglomerates should assume group revenue counts; and the age gate means the model is not intended for under-18 users. If you want genuinely unconditioned weights at this family, that remains Glimmer 30B's Apache-2.0 territory.
What it costs to run
| Lane | Price | Notes |
|---|---|---|
| Downloads / local weights | Free | Hugging Face, AWS SageMaker, Google Cloud Model Garden, Microsoft Foundry |
| Hosted API (7 platforms: OpenRouter, Baseten, Fireworks, Groq, Modal, Nebius, Together AI) | $1.25 in / $4.25 out per M; ~88% cache discount (≈$0.15/M cache hit) | 1M context; tool calls; OpenAI-compatible on Baseten/Together |
| Artificial Analysis cost per task | $0.84 per AA-task blend | 131 tok/s output; ~110M reasoning tokens per full index run |
| Contributor tier | Not offered | The $0.10/$0.20 Contributor pricing exists for closed Spark 1.2/1.3, not Spark Open |
- The catch: Artificial Analysis's own summary puts Spark Open among the smarter models but "particularly expensive when comparing to other open-weight models of similar size" – hosted, you pay Spark-1.3 pricing for an index score nine points lower.
- Concrete: at full rates a long agentic loop bill is roughly $12.50 per 10M input tokens and $85.00 per 20M output tokens – before the ~88% cache haircut softens it. Price a busy agent stack against Glimmer 30B on a single GPU with our API cost calculator.
Can you run it locally? Yes – on a very specific desk
Meta shipped its launch playing nicely with Apple: the reference local setup is two Mac Studios with M5 Ultra 512GB (each pair linked by Thunderbolt 5 – Meta recommends two cables per pair) or four Mac Studios at 256GB each (every Mac connected to every other). The enablers are RDMA-over-Thunderbolt (macOS 26.2+) and MLX 0.30.1+. Apple measured 36.7 tokens per second on the four-Mac tensor setup – comfortably faster than reading speed – with peak power around 300W per Mac at load.
| Route | What it takes | Speed (Apple/Meta-cited) |
|---|---|---|
| 2 × Mac Studio M5 Ultra 512GB (MLX) | 512GB tier ships late October 2026 – price unpublished; two TB5 cables per pair recommended | ~15.7 (pipeline) to 25.0 (tensor) tok/s during Apple's tests |
| 4 × Mac Studio M5 Ultra 256GB (MLX) | ≈$43,196 before tax in the US today; full-mesh TB5 cabling; ~4-hour Meta setup guide | 36.7 tok/s measured |
| One B200 node, NVFP4 (vLLM/TensorRT-LLM) | 8 × GPUs, ≥700GB aggregate VRAM everywhere; Tensor + Expert Parallel (TP8 + EP, 32 experts per GPU) | Hosted-provider lane; 131 tok/s like-for-like class |
| BF16 full precision | ~2TB VRAM – a 4-node × 8 × H200 cluster (TP=32) in the shared recipe | Training/fine-tune lane (TorchTitan supported) |
| Community quants (llama.cpp GGUF, MLX 3-bit/4-bit) | Day-one support; Unsloth/others republishing | Yours to measure – expect memory-bandwidth-bound decode |
One launch-week wrinkle: the 512GB Mac Studio – the config the two-Mac headline setup needs – was not orderable when Spark Open landed. Apple says late October; four 256GB machines (≈$43.2k) were the only Mac route on day one. And there's a speculator: DSpark, a trained draft model from Inferact, cuts latency by proposing 7-token blocks with adaptive verification – it ships on by default for NVFP4 in vLLM. Also notable: SGLang and vLLM ship tuned Helion/Triton kernels for the MoE routing, giving 1.03–3.95× kernel-level speedups on NVIDIA and AMD cards per the PyTorch team's post.
And the brutal context from August's coverage: even Apple's flagship 512GB machine can't hold the field's largest open model in one box – Kimi K3 (2.8T total) needs about 594GB at its most aggressive quantisation. Spark Open's 493GB at 4-bit needs a pair. The desktop caught up to frontier open weights; the frontier did not hold still.
The India angle
- The local-AI rung for India is not this model. Four 256GB Mac Studios and a full TB5 mesh – and only late-October 512GB units – are overkill for desktop users. For real consumer budgets, our GPU guide runs Glimmer 30B at ~17GB of RAM on cards well under ₹50k.
- Rough landed cost of the reference route: four 256GB Mac Studios ≈ $43,196 stateside (≈ ₹45–50 lakh with India GST + duty at prevailing rates). The 2 × 512GB config awaits Apple India's pricing for the late-October tier. Power: about 300W per Mac at load.
- The practical Indian use-case is API-first: the $1.25/$4.25 hosted lane via OpenRouter or the other six platforms, with privacy/compliance-conscious buyers picking the self-host option only when a data-guarantee clause demands it.
Which to pick
- Pick Spark Open if: you need a downloadable frontier-class agentic model with US origin, audited standalone safety (compliance logos matter), 1M context, and you can fund a 2-4 Mac Studio/B200-node class cluster or a hosted lane at Spark pricing.
- Skip if: raw capability is the only metric (GLM 5.3 and Kimi K3 lead most raw rows and scoring), you run on one consumer GPU (that's Glimmer 30B), or you want floor-cheap hosted tokens (the closed Spark 1.3 has the Contributor tier; Spark Open does not).
- Consider instead: Muse Wonder for the closed frontier line (a higher AA score at $4/$20 API pricing), Muse the agent for the app experience, and local vs cloud AI for the framework decision.
What this means if you run local AI
Two things actually changed for the local-AI desk. First, the ceiling moved: open weights around 800B-class sparse MoE now have first-class recipes for desktop clusters, not just datacentres – MLX's RDMA-over-Thunderbolt story is the beginning of a real trend, with Apple shipping the tools and manufacturers tuning for it. Second, 42B-active sparse architectures with 1M context are a template we expect to see trickled down – Glimmer 30B already distilled from this lineage, and our VRAM calculator is the place to sanity-check what a future distilled rung will need.
What didn't change: nobody needs Spark Open at home. It is a datacentre-and-desktop-cluster model for teams who need the licence and safety profile – hobbyists get far more value from the starter path and Glimmer-class models. If your reason to care is "open," your reason to stay on Glimmer/Kimi/GLM quants is unchanged today.
Bottom line
Muse Spark Open matters less for what it scores and more for what it ships with: the first genuinely safety-audited open-weights frontier agentic model, a licence aimed squarely at rescuing enterprise buyers from both Apache ambiguity and platform lock-in, and day-one tooling across every GPU vendor that counts – including a Mac Studio path that actually exists. The capability gap to China's open flagships is real, and the closed Spark 1.3 remains smarter per index point. But in the lane that matters to people who need to ship agentic AI into production while keeping the weights in hand – the rung disappears: there was no comparable option two weeks ago, and now there is one, on a free licence, with a printed compliance story.
Sources
- Meta AI Research – "Introducing Muse Spark Open" (15 Oct 2026)
- Hugging Face – Muse Spark Open model card (incl. licence text and comparison table)
- Apple Machine Learning Research – MLX distributed inference measurements (15 Oct 2026)
- PyTorch – Muse Spark Open kernel optimisations (15 Oct 2026)
- Artificial Analysis – Muse Spark Open profile (independent index, speed, cost)
- Inside Cupertino – Mac Studio configuration and pricing reporting
- OpenRouter – hosted pricing and context window listing
FAQ
What is Muse Spark Open?
An open-weights model Meta released on October 15, 2026: 816B total parameters (42B active per token, sparse MoE), a 1M-token context, native text/image/audio/video input, and text output. It shares one base model with the closed Muse Spark 1.3 and adds safety post-training so it can run standalone without Meta's system guardrails. Weights are free to download under the Muse Model License.
Is it open source?
No - it is open-weights software under a custom licence. The Muse Model License permits free commercial use, fine-tuning, distillation and redistribution, but requires a "Built with Muse Spark Open" notice and a separate agreement with Meta only for Model-as-a-Service businesses over $20 million aggregate gross revenue in any 12 months. It is not OSI-approved open source; Apache-2.0 remains available on the smaller Muse Glimmer 30B.
How does it compare to GLM 5.3, Kimi K3 and Qwen3.8?
On Meta's own model card, GLM 5.3 scores 45 on the Artificial Analysis Intelligence Index vs 39 for Spark Open (Kimi K3: 44, Qwen3.8-2.4T: 40) - the rivals also sweep most raw coding rows. Spark Open instead tops the agentic rows (JobBench, MCP-Atlas tool calling) and dominates safety testing: 0.5 vs 62.9-86.4 on StrongREJECT v2 adversarial jailbreaks. Meta's "most capable US open-weight" claim is scoped to US origin, not a global #1.
How much does it cost?
The weights are free. Hosted inference runs $1.25 per million input tokens and $4.25 per million output tokens (with about 88% caching discount) on OpenRouter, Baseten, Fireworks, Groq, Modal, Nebius and Together AI. Artificial Analysis benchmarks the blend at about $0.84 per index task at 131 output tokens per second. Unlike closed Spark 1.2/1.3, no cheap Contributor mode is offered for Spark Open.
Can I run it locally?
Yes - over multiple machines. Apple's supported setups: two Mac Studios with M5 Ultra 512GB or four with 256GB each, linked by RDMA-over-Thunderbolt 5, running MLX 0.30.1+ - the 4-bit runtime footprint is about 493GB, so no single desktop holds it (the NVFP4 file on disk is ~537GB; BF16 over 1.6TB). Apple measured 36.7 tokens per second on the four-Mac setup, about $43,200 before tax. For a single desktop, use Muse Glimmer 30B (~17GB at 4-bit) instead.
Should I use Spark Open or Muse Glimmer 30B for local work?
Glimmer 30B - it is Apache-2.0 licensed, runs on consumer rigs and matches common local-AI tasks. Use Spark Open when you need its scale, safety evaluation record, 1M context and US-origin licensing in an enterprise deployment: it is built to serve agents across clusters, not to run solo on one gaming machine.