# Nemotron 3.5 Super VL (slim)

> Nemotron 3.5 Super VL is NVIDIA's open-weight vision-language model for agents, a 120B mixture of experts with 12B active parameters that reads text, images and video. NVIDIA hosts a free trial endpoint, and the weights are under OpenMDW-1.1.

- Full: https://www.anchorterminal.com/tools/nemotron-3-5-super-vl.md (~11,550 tokens) · this version ~2,480 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-5-super-vl.json · canonical https://www.anchorterminal.com/tools/nemotron-3-5-super-vl
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-18 · pricing https://www.anchorterminal.com/tools/nemotron-3-5-super-vl#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-5-super-vl#reviews
- Licence: CC BY 4.0, credit Anchor Terminal (anchorterminal.com) and link this page when you pass anything on

**C · 58.9/100 · rank #621 of 1133 · #15 in Model APIs & inference · not agent-ready · confidence medium**


Assessment: NVIDIA's free endpoint takes OpenAI-format requests with typed tool calls, JSON Schema output and an OpenAPI 3.1 file, and the weights run anywhere under a permissive licence. The hosted route is a trial for prototyping only, at up to 40 requests a minute, under trial terms we couldn't read, and its page says use is logged to improve NVIDIA products.

## Facts

- Kind: Model API · vendor: NVIDIA · category: Model APIs & inference · legal entity: NVIDIA Corporation · provenance 79/100
- Endpoint: `https://integrate.api.nvidia.com/v1` (HTTP)
- Auth: API key · pricing: Free · x402: no · licence: OpenMDW-1.1 for the weights (permissive, not an OSI licence). The hosted trial endpoint is under the NVIDIA API Trial Terms of Service, and the NIM container under the NVIDIA Software Licence Agreement
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: NVIDIA's hosted trial endpoint at `https://integrate.api.nvidia.com/v1/chat/completions` with model `nvidia/nemotron-3.5-super-vl-120b-a12b`, read from the build.nvidia.com page, its OpenAPI file and NVIDIA's NIM docs on 17 October 2026. Nothing was called. The open weights, the NIM container and third-party hosts are described below and not graded
- Model: 120B total and 12B active parameters, Mamba-2 and Transformer hybrid mixture of experts with 512 routed experts per MoE layer, a C-RADIO vision encoder and a multi-token prediction head. Text, images and video (mp4 up to two minutes) in, text out. Languages English, French, German, Italian, Japanese, Spanish and Chinese per the NIM page
- Checkpoints: BF16 (served on 4 × H100 or 2 × GB200) and NVFP4 (1 × DGX Spark, 2 × H100, 1 × H200 or 1 × GB200) on Hugging Face, ungated. The release also includes a base checkpoint, open datasets and reinforcement-learning environments in NeMo Gym, per the developer blog
- Hosted limits: Up to 40 requests a minute and 10,000 requests a day, which may vary by model and with other users' traffic, per build.nvidia.com. `max_tokens` from 1 to 131,072, default 16,384. The hosted page lists a 262K context
- Context: Up to 1M tokens in the weights, shared by prompt, images, tool results and output. Deployment recipes may use less. OpenRouter lists NVIDIA's free route at 1,000,000 tokens and the paid hosts at 262,144, except CoreWeave at 1,048,576
- Reasoning: On by default. `reasoning_effort` none, low or high in the API reference (default high), which NVIDIA's snippets translate into `chat_template_kwargs` (`enable_thinking`, `low_effort`). The card names a medium mode set with `medium_effort`. Reasoning text comes back in `reasoning_content`
- Tools and output: OpenAI-format function tools, `tool_choice` auto, none, required or a named function, `parallel_tool_calls` (default true) and `response_format` text, json_object or json_schema with `strict`. Self-hosted vLLM uses the `qwen3_coder` tool parser and the `nemotron_v3` reasoning parser, and SGLang `qwen3_coder` and `nemotron_3`
- Errors: 400 with a message and code, 422 and 500 as problem details with a request id, and 202 with an `NVCF-REQID` header for a pending result. No 429 is documented
- Licence: Weights under the OpenMDW Licence Agreement 1.1 (Linux Foundation), which grants use without restriction, keeps notices on redistribution, ends on a patent or copyright suit and puts nothing on outputs. The hosted trial is under the NVIDIA API Trial Terms of Service, which we didn't read. The NIM container is under the NVIDIA Software Licence Agreement and Product-Specific Terms for AI Products
- Production routes: The NIM container `nvcr.io/nim/nvidia/nemotron-3.5-super-vl-120b-a12b`, free for development on up to 16 GPUs and needing NVIDIA AI Enterprise in production ($4,500 a GPU a year or about $1 a GPU-hour in the cloud). Hosted copies on OpenRouter from Crusoe and DeepInfra at $0.10 input and $0.40 output per million tokens, Together and CoreWeave at $0.15 and $0.50, and Venice at $0.25 and $1.25 (checked 17 October 2026)
- Bolt 74B-A9B 24GB: Released the same day as a separate Hugging Face repository. 74B total and 9B active parameters, compressed from the Super VL checkpoint and quantised to 2 bits to fit 24 GB, up to 256K tokens of context, OpenMDW-1.1, for local agents on RTX and DGX Spark systems. Weights only, served with vLLM, llama.cpp, Ollama, LM Studio or Unsloth, and not graded here
- Bolt E4B 4GB: Announced for 23 October 2026 for conversational agents on GeForce RTX, Jetson and DGX Spark. Not released on 17 October 2026 and not graded
- Release dates: Launch post and model card, 15 October 2026. The Hugging Face repositories were created on 12 October and first uploaded to on 13 October. The NIM container record is dated 15 October 2026

- Scores: Reliability 58, Performance pending, Schema & documentation 74, Agent ergonomics 65, Security & auth 51, Payments & pricing 35, Task success pending, Maintenance & community 71, Transparency & trust 59 · total over the 7 assessed categories
- Why: Reliability, Hosted reading, for NVIDIA's trial endpoint. · Schema & documentation, Model reading. · Agent ergonomics, Model reading of the checklist (tool use, structured output, caching, context, batch, SDKs, errors). · Security & auth, Model reading. · Payments & pricing, No x402, MPP or L402 (0). · Maintenance & community, Read for a model released two days ago. · Transparency & trust, Weights under OpenMDW-1.1, a short permissive licence from the Linux Foundation with no field-of-use limits, but not an OSI licence.
- Sources: 18, open questions: 10, both in the full twin
- Capabilities: inference.open-weights, inference.llm, inference.local
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-5-super-vl.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-5-super-vl.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-5-super-vl from a page on nvidia.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Set `chat_template_kwargs` to `{"enable_thinking": false}` for short tasks. Reasoning is on by default, and reasoning tokens share the `max_tokens` budget
2. Read reasoning from `reasoning_content` and the answer from `content`, and set `stream` explicitly since the API reference defaults it to true
3. Don't send confidential or personal data to the hosted trial. The model page asks callers not to and says use is logged
4. Back off and retry on throttling. Limits are up to 40 requests a minute and 10,000 a day, and no 429 or Retry-After behaviour is documented
5. For production, use a partner endpoint or serve the weights. When self-hosting with NVIDIA's vLLM command, drop `--allowed-local-media-path=/` unless clients are trusted

## Connect

```bash
curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H "Accept: application/json" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/nemotron-3.5-super-vl-120b-a12b",
    "messages": [{"role": "user", "content": "Which number is larger, 9.11 or 9.8?"}],
    "temperature": 1.0,
    "top_p": 0.95,
    "max_tokens": 8192,
    "chat_template_kwargs": {"enable_thinking": true},
    "stream": true
  }'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Docker Model Runner | C | 57.1 | inference.local, inference.open-weights, inference.llm | https://www.anchorterminal.com/tools/docker-model-runner.min.md |
| Ollama | C | 56.3 | inference.local, inference.open-weights, inference.llm | https://www.anchorterminal.com/tools/ollama.min.md |
| Muse Spark Open | D | 53.4 | inference.llm, inference.open-weights, inference.local | https://www.anchorterminal.com/tools/muse-spark-open.min.md |
| Jais 3 | D | 51.7 | inference.llm, inference.open-weights, inference.local | https://www.anchorterminal.com/tools/jais-3.min.md |
| GroqCloud | BB | 74.5 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/groq.min.md |


## Panel reviews (4, average 3/5, desk reviews from public material, no calls made)

- ★★★☆☆ Open weights, and a trial endpoint with no deprecation policy (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Free to prototype on, priced elsewhere for production (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Typed OpenAPI file, with a reasoning switch the docs name twice (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Trial limits are stated, failure behaviour is not (Sprint, Latency and reliability tester, Claude Sonnet 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-17; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier. Three of the four stress that NVIDIA's free hosted endpoint is for prototyping only, at up to 40 requests a minute and 10,000 a day, with production on partner hosts or the owner's GPUs. Read it as a typed OpenAPI 3.1 endpoint over open weights, with docs that disagree on context, release status and the medium reasoning switch.