Best Hosts for Muse Spark Open and Nemotron 3.5 Super VL
The short answer
For Muse Spark Open every host charges the same $1.25 / $4.25 per million tokens, so pick on features. For Nemotron 3.5 Super VL, Crusoe and DeepInfra are cheapest at $0.10 / $0.40, and CoreWeave is the one to use if you need the full 1M-token context. Both models were released as open weights on October 15, 2026. Prices and limits below are the per-provider listings on OpenRouter’s endpoint API, read October 19, 2026; check a provider’s own page before you commit volume.
Muse Spark Open (Meta, 816B / 42B active)
| Host | Input / 1M | Cached input / 1M | Output / 1M | Context | Quantization listed |
|---|---|---|---|---|---|
| Together AI | $1.25 | $0.15 | $4.25 | 1,048,576 | not stated |
| Fireworks | $1.25 | $0.15 | $4.25 | 1,048,576 | not stated |
| Baseten | $1.25 | $0.15 | $4.25 | 1,048,576 | FP4 |
| Modal | $1.25 | $0.15 | $4.25 | 1,000,000 | not stated |
| Nebius | $1.25 | — (none listed) | $4.25 | 1,048,576 | FP4 |
That is the same rate Meta charges for its closed Muse Spark 1.3 on the Meta Model API. Together’s launch promotion (free until October 18, 2026) has ended.
How to choose: with price flat, the decision is about the workload around the model. Agents that resend a long system prompt benefit from the $0.15 cached-input rate on Together, Fireworks, Baseten and Modal — Nebius lists none. If you expect to fine-tune or distil (the Muse Model License allows both), start on Baseten or Together, where a dedicated deployment of your variant is a configuration change rather than a migration. Remember the licence terms: products must display “Built with Muse Spark Open”, and a company reselling the model as an API with over $20 million in revenue needs a separate agreement with Meta.
Nemotron 3.5 Super VL (NVIDIA, 120B / 12B active)
| Host | Input / 1M | Cached input / 1M | Output / 1M | Context | Max output | Quantization |
|---|---|---|---|---|---|---|
| Crusoe (OpenRouter listing) | $0.10 | $0.08 | $0.40 | 262,144 | 235,929 | BF16 |
| DeepInfra | $0.10 | $0.08 | $0.40 | 262,144 | 235,929 | FP4 |
| Together AI | $0.15 | $0.02 | $0.50 | 262,144 | 235,929 | not stated |
| CoreWeave | $0.15 | $0.02 | $0.50 | 1,048,576 | 943,718 | NVFP4 |
| Venice | $0.25 | $0.125 | $1.25 | 262,144 | 32,768 | FP4 |
| Baseten | launch partner (own Model API) |
How to choose:
- Cheapest per token: Crusoe if you want BF16 (unquantized) weights, DeepInfra for FP4 — same price.
- Heavy prompt caching: Together or CoreWeave. Their cached input is $0.02 against $0.08 at Crusoe and DeepInfra, so an agent that reuses most of a long prompt can end up cheaper on the “more expensive” hosts.
- Long documents or video beyond 262K tokens: CoreWeave is the only listed host serving the full ~1M-token window.
- Throughput: NVIDIA’s launch post shows Nemotron 3.5 Super VL at over 600 tokens per second on Artificial Analysis agentic tasks on Baseten’s endpoint.
- Avoid for long outputs: Venice caps output at 32,768 tokens and costs the most.
CoreWeave showed 96.4% recent uptime on OpenRouter when we checked, against 100% for the others; route through two hosts if this model is on your critical path.
Self-hosting instead
- Nemotron 3.5 Super VL is practical to run yourself: NVIDIA publishes base, post-trained and NVFP4 checkpoints on Hugging Face under the OpenMDW licence, optimised it for a single DGX Spark, and documents Dynamo recipes for multi-GPU serving. See best LLM inference engines for the serving stack.
- Muse Spark Open needs roughly 700 GB of GPU memory in NVFP4 (2 TB in BF16) per Meta’s model card — an 8-GPU server. At $1.25 / $4.25 the hosted APIs win until volume is very large.
Background on each model: what is Muse Spark Open and what is Nemotron 3.5 Super VL. To compare against closed models, use current API prices and the cheapest inference providers for open-weight models.
Last verified: October 19, 2026. Per-host prices, context and quantization from OpenRouter’s public endpoint listings; uptime is OpenRouter’s recent measurement and changes hourly.