Meta's Muse Spark Open: 816B AI Model You Can Download Free

Quick answer

Meta released Muse Spark Open on October 15, 2026: the open-weights version of its Muse Spark model — 816 billion parameters (42 billion active per token), a 1-million-token context window, and native multimodal input, free to download from Hugging Face. The real headline is hardware: Meta optimized it with Apple's MLX framework to run across two M5 Ultra Mac Studios. The real fine print is the license: open weights, but not open source — you must display "Built with Muse Spark Open," and API hosts with over $20 million in gross revenue need a separate written agreement with Meta first. Hosted API pricing is listed at $1.25 per million input tokens, while independent benchmarks rank the model behind China's Kimi K3 and GLM-5.3 on overall capability.

What Meta actually released

Six days ago, Meta put the weights of its biggest open model on the table. Muse Spark Open is the open-weights sibling of Muse Spark, the agentic model Meta has been pushing through its Muse Code developer tool and the Meta Model API. The numbers, from Meta's own model card: 816 billion total parameters in a mixture-of-experts design, with roughly 42 billion active for any given token, a 1-million-token context window, and native multimodal handling — text, images, video and audio in, text out, across more than 100 languages. Meta says it trained the model for long-horizon agentic work: coding, computer use, and multi-step tool calling.

The weights ship in two flavors on Hugging Face — full-precision BF16 and quantized NVFP4 — and Meta also lists them on Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. One catch before you click download: the Hugging Face page gates access behind a form asking for your legal name, date of birth, country and affiliation, plus manual approval. This is not a click-and-clone Llama release.

Meta calls it "the most capable US open-weight model to date." That sentence has two load-bearing words: "US." On the independent Artificial Analysis Intelligence Index, Muse Spark Open scores 39 — tied for eighth place on the open-weight leaderboard with DeepSeek V4.1 Flash, and behind seven Chinese models: Xiaomi's MiMo-V2.6-Pro (46), Z.AI's GLM-5.3 (45), StepFun's Step 5 Preview (44), Moonshot's Kimi K3 (44), and Alibaba's Qwen3.8 (40). It's a genuine milestone for American open weights; it is not the best open model in the world, and Meta's own phrasing quietly concedes that.

The headline is two Mac Studios

Meta didn't just release weights. It released a reference story about running an 816-billion-parameter model on a desk. With Apple's MLX framework, Meta says two Mac Studios with M5 Ultra chips "behave like one supercomputer with a terabyte of unified memory." Apple's published numbers for the setup: about 25 tokens per second across two machines using tensor parallelism over Thunderbolt, and more than 36 tokens per second with four.

Here's what the press releases underplay. The 512GB configuration Meta's demo depends on isn't on sale yet — Apple is expected to ship it later this month — and a 4-bit quantized version needs about 493GB for the weights alone, so it doesn't fit on a single 512GB machine once inference memory is added. The current 256GB Mac Studio is already priced above $10,000, per MacRumors reporting, with the M5 Ultra starting around $5,499. Two top-spec machines just to host one model locally is "desktop AI" only in the sense that a server rack on your desk is still a desk. This is a demo for labs and well-funded shops, not a new normal.

How good is it? Read the table, then the fine print

Meta's published scorecard says Muse Spark Open leads on 6 of 11 open-weight benchmarks, with Kimi K3 leading on 4. The standout is MCP-Atlas at 89.2 — a test of agent tool use where it posts the highest score in Meta's own comparison table. That lines up with what the model is for: not raw chat, but agentic work where calling tools correctly matters more than sounding clever.

The other side of the ledger: on pure coding benchmarks like Terminal-Bench 2.1, third-party write-ups note it trails Kimi K3 and GLM-5.3. The Artificial Analysis Index (39) puts it below Qwen3.8 (40) and level with Europe's Mistral Large 4 at 38. And every score here comes from Meta's own published tables or vendor-reported figures — independent replication of the speed and quality claims is still pending, which is exactly where you should be skeptical. If you're comparing open models for an agentic workload, test tool-calling accuracy on your own tasks; the leaderboard won't settle this for you.

The license: open weights, not open source

This is the part that will matter most in the long run, and the part most coverage rushed past. Muse Spark Open ships under the Muse Model License — a Meta-authored license, not Apache 2.0, not MIT, and not the Llama Community License. It is not open source in the OSI sense. What it grants: use, modification, fine-tuning, distillation, distribution, and selling products built on the model, free of charge. What it demands:

• Every product built on it must prominently display "Built with Muse Spark Open" in its interface, documentation or settings.
• If you sell model access as a service — an API or similar — and your company plus affiliates exceed $20 million in gross revenue over any 12-month period, you must notify Meta and sign a separate written agreement before offering the model. Internal use, research, and embedding the model inside your own app's features are exempt; reselling inference at scale is not.
• An attached usage policy bans malware, weapons development, fraud, impersonation and use by under-18s, and Meta added extra safety post-training for this open release.

The shape is deliberate. Meta moved the commercial carve-out from who you are (Llama's 700-million-user threshold) to what you sell (inference as a service, above a revenue line). Most developers — building apps, fine-tuning, deploying internally — pay nothing and sign nothing. If your business is selling API access to the model, read the license with a lawyer before you ship.

Three ways to use it, with real costs

A worked cost example at the Artificial Analysis-listed rates: a coding session consuming 2 million input tokens and 200,000 output tokens costs 2 × $1.25 + 0.2 × $4.25 = $3.35. Provider pricing varies, so confirm the rate on the platform you actually use — $1.25/$4.25 is the reference, not a universal price.

Context on where that price sits: it's the same $2/$10 mid-tier pricing cluster Anthropic, OpenAI and Google have all converged on (see our breakdown of Claude Haiku 5.5's pricing shake-up), but one tier cheaper on output — positioned to win developers building token-hungry agents.

Pros and cons

What's strong

• Best-in-class open-weight tool use (MCP-Atlas 89.2) — the metric that actually matters for agents.
• 1-million-token context with native multimodal input, genuinely useful for long-horizon coding and document work.
• Free weights under a license that most developers — internal use, apps, fine-tuning — will never bump against.
• One of the cheapest open-weight API tiers for agentic workloads at the listed $1.25/$4.25 rates.

What's not

• Not open source — attribution requirement and the $20M API-host clause will bother purists and resellers alike.
• Trails Chinese rivals on pure coding benchmarks and overall intelligence scores; "most capable US open-weight" is a qualified crown.
• The Mac Studio story needs hardware that isn't shipping yet — 25 tokens/sec across two machines is a demo, not a deployment strategy.
• All headline benchmarks are vendor-reported; independent verification is still thin.

Who it's for — and who should skip it

Look at it now if you're building tool-heavy agents that need to stay on your own infrastructure, or if your organization has a policy against Chinese-built models and you've been waiting for a US open-weight option in this range. Skip it for now if you mainly want a coding agent — Kimi K3 and GLM-5.3 score better today — or if you're expecting a laptop-friendly local model: the smallest community-quantized builds still need hundreds of gigabytes of memory, and Meta's own local-tier answer is the 30B Muse Glimmer under Apache 2.0, not this.

FAQ

Is Muse Spark Open really free?

The weights are free to download from Hugging Face under the Muse Model License, and you can fine-tune and sell products built on it with "Built with Muse Spark Open" attribution. But running it yourself needs hundreds of gigabytes of GPU memory, so most people will pay the practical cost: hosted API access at around $1.25 per million input tokens.

Can I run Muse Spark Open on my laptop?

No. The NVFP4 checkpoint needs about 700GB of GPU memory; the full BF16 version needs around 2TB. A normal laptop or gaming PC is far short. Meta's 30B Muse Glimmer is the Meta model actually aimed at local use.

Is Muse Spark Open open source?

Not in the strict sense. It's "open weights" under Meta's custom license: free to download, fine-tune and commercialize, but with an attribution requirement, an acceptable use policy, and a separate agreement for API businesses above $20 million in revenue. No OSI-approved open-source license applies.

How does it compare to GLM-5.3 and Kimi K3?

On the Artificial Analysis Intelligence Index, GLM-5.3 scores 45 and Kimi K3 44 against Muse Spark Open's 39. Meta's own table shows Muse Spark Open leading on tool use (MCP-Atlas 89.2) but trailing both on pure coding tests. It's an agent model first, a coding model second — and both Chinese rivals remain the stronger all-rounders for now.

Key takeaways

• Meta released Muse Spark Open on October 15, 2026: 816B parameters (42B active), 1M-token context, free weights on Hugging Face.
• The two-Mac-Studio demo needs hardware that isn't on sale yet — treat it as a lab showcase, not a buying guide.
• The license is open weights, not open source: "Built with Muse Spark Open" attribution, and a $20M revenue threshold for API resellers.
• Best-in-class open-weight tool use, but independent benchmarks rank it behind GLM-5.3, Kimi K3 and MiMo-V2.6-Pro overall.
• For most developers, the hosted API (~$1.25/$4.25 per million tokens) is the only practical way in — the download is for researchers and infra teams.

Six days after release, Muse Spark Open is exactly what Meta promised — an American open-weight model with real agentic muscle — wrapped in exactly the license and hardware story you'd expect from Meta. The weights are the gift; the fine print is the price. Read both.