Nemotron 3.5 Super VL
工具调用推理缓存结构化输出
最低输入价
$0.09
Io Net
上下文
1M
比较服务商
每个点是一家服务商。从左到右是每百万 token 的输入价,从下到上是输出价。
基础价格
9最新价格检查于 Nov 18, 2026。
Nemotron 3.5 Super VL · 已公布价格
Nemotron 3.5 Super VL · 权重许可证
| 模型权重 | 协议 |
|---|---|
| 开放权重 | OpenMDW 1.1 |
Nemotron 3.5 Super VL · 权重与量化
| Item | Configuration / evidence |
|---|---|
| Checkpoint | Nemotron 3.5 Super VL 120B-A12B NVFP4; OpenMDW 1.1 |
| Quantization | W4A4 on Blackwell; W4A16 on non-Blackwell, per serving configuration |
Nemotron 3.5 Super VL · 自部署硬件
| Item | Configuration / evidence |
|---|---|
| Supported hardware | Blackwell GB200 / DGX Spark; Hopper H100 / H200 |
Nemotron 3.5 Super VL · 部署方式
| Item | Configuration / evidence |
|---|---|
| Serving engines | vLLM / SGLang; NVIDIA GPU-accelerated systems |
Nemotron 3.5 Super VL · 上下文与渠道限制
| Item | Configuration / evidence |
|---|---|
| Local context | Up to 1M; recipes may reduce context for GPU memory and concurrency |
Nemotron 3.5 Super VL · 已知风险与限制
| Item | Configuration / evidence |
|---|---|
| Benchmark scope | NVIDIA evaluations use NeMo Gym recipes; not independent LMSpeed measurements |
| Model identity | This is Super VL 120B-A12B; no BOLT variant or edge latency result verified |
Nemotron 3.5 Super VL · 使用场景
| Item | Configuration / evidence |
|---|---|
| Enterprise use cases | Document and video understanding; coding, reasoning and tool-using agents |
上下文与限制
It does not publish a paid per-token tariff.
[2][3]NVIDIA Nemotron 3.5 Super VL has 120 billion total parameters and 12 billion active parameters.
[2]The model card currently describes the NVFP4 quantized release.
[2]Use of the model is governed by the OpenMDW License Agreement version 1.1.
[2][3]Access to the hosted trial is governed by the NVIDIA API Trial Terms of Service.
[2][3]
关于 Nemotron 3.5 Super VL
OpenRouter lists nvidia/nemotron-3.5-super-vl:free with NVIDIA as the provider and free input/output. NVIDIA Build also describes hosted trial access. Check current trial terms and availability rather than assuming unrestricted production use.
The checked NVIDIA model card provides trial terms but no paid token tariff. Paid hosting offers are provider prices and should be compared separately.
NVIDIA lists 120B total parameters, 12B active parameters, and up to 1M context tokens. The checked Build card describes an NVFP4 quantized release under OpenMDW-1.1.
The NVIDIA model card identifies a model license and a separate hosted trial agreement. Those documents govern different relationships. Downloading and running an artifact requires reviewing the model license and planning the infrastructure used to serve it. Calling the NVIDIA-hosted trial requires reviewing the trial service terms and the account’s access conditions. The public card does not supply a paid NVIDIA token tariff or a universal trial quota, and this guide does not infer either. Keep the exact artifact and the service endpoint in a deployment record. If an OpenRouter free route is used, retain its :free suffix and the linked terms as well. A zero input/output tariff describes the listed route at the checked date; it does not establish unrestricted capacity or remove your own application’s operating costs.
No. NVIDIA’s architecture description identifies the model and its supported input types. It does not give a fixed number of billed tokens per document, screenshot, or video second. Such a conversion needs the chosen endpoint’s documented accounting rules. The 120B total and 12B active parameter figures likewise do not establish a provider bill, GPU rental requirement, or throughput guarantee. For a hosted estimate, begin with actual or explicitly assumed usage categories and the selected provider’s published tariff. For a self-hosted estimate, use a separately sourced infrastructure plan. Keep precision in the record because the checked Build card describes NVFP4, while other routes can have different published configurations. This guide does not convert architecture details into unmeasured speed or quality claims. Missing pricing and usage rules stay unresolved until a suitable source is available.
来源与注脚
- [1]来源核验日期:
- [2]
- [3]
输入、输出和缓存读取的单位是每百万 token 的美元价格。速度是每秒输出 token 数。可用率是来源报告的在线比例。