Meta Muse Spark Open for private enterprise AI: the call

The call

Meta published the weights, hosting partners, and hardware guidance on 15 October (vendor). Coverage the same day: MarkTechPost and Unite.AI (independent).

What it is, and what it is not

What it is. A mixture-of-experts model with 816 billion parameters, about 42 billion active per token, a 1-million-token context, and text, image, video, and audio in, text out (vendor). Weights ship in BF16 at about 1.6 TB and NVFP4 at about 530 GB. Hosted on Baseten, Fireworks AI, Groq in limited preview, Modal, Nebius, OpenRouter, and Together AI, and listed on Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models (vendor). Reasoning depth runs from low to xhigh, and the default is xhigh (independent, MarkTechPost).

What it is not. It is not the best open model. Meta’s own table puts it at 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3 at 45, Kimi K3 at 44, and Qwen3.8 at 40, and behind all three on Terminal-Bench 2.1 (vendor table, via MarkTechPost). It is not independently evaluated yet; Meta says a full evaluation report is coming (independent, Unite.AI). And it is not unconditionally open. The Muse Model License permits use, fine-tuning, and distillation, but products must display “Built with Muse Spark Open”, and model-as-a-service providers above $20 million in 12-month revenue need a separate agreement with Meta. Internal use and your own product features are exempt from that clause (independent, Unite.AI).

The number nobody is quoting

“Runs on two Mac Studios” is the line everyone is repeating. Two 512 GB machines make about a terabyte of memory (vendor). The number that decides the business case is what a hosted call costs against what you already buy.

Muse Spark Open lists at $1.25 input and $4.25 output per million tokens on OpenRouter, Baseten, and Nebius (independent, MarkTechPost). Claude Haiku 5.5 lists at $0.10 and $0.50, Claude Sonnet 5.5 at $2 and $10 (vendor, via BenchLM), and GPT-6 Luna at $0.10 and $0.50 (vendor, via CellCog). Take a service call of 1,000 tokens in and 300 out, no reasoning.

Now add the default. At xhigh, and assuming your host bills reasoning tokens at the output rate, every 1,000 reasoning tokens adds $4.25 per 1,000 calls. At 583 reasoning tokens a call, Muse costs more than Sonnet 5.5 answering without reasoning: ($5.00 − $2.53) ÷ $4.25 per 1,000 tokens (our maths). Leave effort on the default and half-price becomes full price.

Self-hosted, the arithmetic changes shape. NVFP4 weights are 530 GB against Meta’s recommended 700 GB of GPU memory, which leaves 170 GB for the context cache and everything else (our maths). BF16 at 1.6 TB needs the recommended 2 TB. A million-token context on 170 GB of headroom is a claim to test, not to assume.

Missing

What Meta has not told an enterprise buyer, as of 16 October:

• An independent evaluation. Every benchmark on the page is Meta’s.
• The full evaluation and safety report, promised “soon”.
• Reasoning tokens per call at each effort level.
• Data retention terms for each hosting partner. Only Together AI states an SLA, at 99.9% (vendor).
• Amazon Bedrock or Vertex AI managed endpoints; Meta names SageMaker, Model Garden, and Foundry Models.
• Context-length throughput on the recommended 700 GB, and on two Mac Studios.
• Whether an internal-only tool must show the “Built with” notice; the licence ties it to products.

Who owns what

Shipping the weights is Meta’s job. Running them safely is ours.

Self-hosted means your team owns patching, evaluation, red-teaming, and the safety post-training Meta warns can degrade once you fine-tune (independent, Unite.AI). From Salesforce, the call is an Apex callout through a Named Credential to your own endpoint, in your VPC or on SageMaker or Foundry. Long xhigh calls belong behind a Platform Event or a queueable, not a synchronous Flow step. Keep the model name and effort level in Custom Metadata so a switch is configuration.

Commercial decision: privacy is worth paying 10x per call for some workloads and none for others. Price the workload, not the model.

Monday morning

1. Step 1. Decide if you need private weights at all. List the workloads where data cannot leave your boundary. If there are none, stop here. Owner: CIO.
2. Step 2. Read the licence. Attribution, the $20 million clause, and the usage policy, against how you would deploy. Owner: legal.
3. Step 3. Shadow-test on 200 real tasks. Use a hosted endpoint whose data terms you have checked, at effort low and xhigh, against your current model; log answers and output tokens per call. Owner: AI platform lead.
4. Step 4. Size the hardware. Quote 700 GB of GPU memory for NVFP4 from your cloud and on-premises suppliers, with the context length you need. Owner: infrastructure lead.
5. Step 5. Hold production for an outside evaluation. Meta’s full report, plus at least one independent benchmark, before any customer-facing use. Owner: security lead.

We move this to SKIP if independent evaluations land well below Meta’s table. We move it to PILOT for everyone if a host prices it near Haiku 5.5.