Nemotron 3.5: Super VL API costs and Bolt local setup

NVIDIA Build supplies a free prototype endpoint and a key-generation flow. Its hosted model ID is nvidia/nemotron-3.5-super-vl-120b-a12b, with a listed 262K context on that route. Use the downloadable model card’s 120B identity; the checkpoint’s 1M maximum is broader than the prototype channel’s current limit.

Vercel AI Gateway offers nvidia/nemotron-3.5-super-vl through several providers. Its provider rows have different rates and context limits. For the example below, install AI SDK 7 or later and set AI_GATEWAY_API_KEY in your environment. Restrict routing to the provider used in a budget when you need that rate and capacity; an order preference alone can still fall back elsewhere.

The table reproduces the specific AI Gateway provider rates checked October 24, 2026, in USD per million tokens. Cache rates apply to reported cache hits. Choose a provider row rather than using the page’s rounded starting-price headline.

For Baseten, 100,000 uncached input tokens plus 10,000 output tokens cost $0.020 + $0.005 = $0.025. The same workload on Together’s listed row is $0.020, with a smaller 262K context. Count the billed reasoning and repeated tool turns in the total output and usage for the task.

Super VL combines Mamba-2, attention and MoE layers with a vision encoder. The October 15 NVFP4 v1.0 GA checkpoint accepts text, images and video and returns text. NVIDIA lists one H200, two H100s, one GB200 or DGX Spark configurations; context, media and concurrency settings belong to each recipe.

Use vLLM 0.32.0 or the card’s exact nightly image, SGLang 0.5.20, or TensorRT-LLM 1.3.0rc28 with its documented patch. Select the command for your hardware, wait for model loading and kernel compilation, then check /v1/models before sending a short request. The full model download, caches and runtime buffers all need space; a repository’s download size is not a peak VRAM measurement.

This section covers Bolt 74B/A9B. Bolt E4B is a different text-only model with up to 64K context, so these projector and allocation recipes apply only to the 74B checkpoint.

Bolt is a distinct 74B-total, 9B-active vision-language model with a 256K maximum context. Its 24GB GGUF variant needs the main checkpoint, vision projector and a separate DFlash drafter for the supplied recipes. The model card recommends more device memory than the variant’s name alone suggests.

The 28GB allocation recipe uses one concurrent sequence, a 65,536-token context and up to six draft tokens. NVIDIA recommends a 32GB-or-larger discrete GPU or a 64GB Spark system for it. The 32GB recipe uses a 262,144-token context and up to seven draft tokens, with a 48GB-or-larger GPU or 128GB Spark system. Follow the current llama.cpp build and the recipe’s cache settings together.

Both official checkpoints use OpenMDW License Agreement 1.1, which permits use, modification and redistribution for commercial and non-commercial work under its conditions. Retain the license and required notices when distributing covered materials. The serving runtime’s own license is a separate part of the stack.

For an API trial, Super VL has explicit hosted routes and provider prices. For a compact local agent, Bolt’s GGUF guide supplies the projector, drafter and memory configuration as a complete path. Compare an accepted task at the intended context and input types, including time spent on reasoning and tools.

Step 5 Preview and MiMo provide other large agent-model routes. For a text-only local agent, LFM2.5 or the Qwen guides may fit a different hardware budget. Use their documented configurations instead of transferring Bolt’s allocation or Super VL’s 1M context to another checkpoint.