Nemotron 3.5 Super VL

# Fix list: Nemotron 3.5 Super VL

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/nemotron-3-5-super-vl, the October 2026 research run, assessed 17 October 2026. Grade C, 58.9 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Nemotron 3.5 Super VL: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Security & auth, 51 out of 100, up to 8.6 more on the total

Why it scored 51: Model reading. `nvapi-` bearer keys sent in a header. NVIDIA's NGC docs say personal keys are limited to the services and role chosen at creation and take an expiry, and service keys can be scoped to actions (25 of 30). Inference only, with no write operation, but no zero-retention option for the hosted trial was found, and the model page asks callers not to send confidential or personal data (8 of 20). The model reads web pages, PDFs and images an agent passes it, and NVIDIA publishes safety, bias and privacy subcards, but no prompt-injection guidance or results were found (5 of 15). No per-call log for API callers was found, only rate limits shown in the account (3 of 15). NVIDIA PSIRT takes vulnerability reports, linked from the card. www.nvidia.com/.well-known/security.txt answered 403 and build.nvidia.com has none, and no certification for the API catalogue was found (10 of 20). 51.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 2. Reliability, 58 out of 100, up to 8.4 more on the total

Why it scored 58: Hosted reading, for NVIDIA's trial endpoint. status.ngc.nvidia.com is a Statuspage site with NVIDIA Build and Cloud Functions components and a readable history (20). In the 90 days to 17 October two incidents, both minor for an API caller. SMS verification for NVIDIA Build was unavailable in China from 23 to 29 September, which affects sign-up, and on 10 August new Cloud Functions deployments failed for about three hours while existing deployments kept running (20). build.nvidia.com gives limits of up to 40 requests a minute and 10,000 a day, and says they vary by model and with other users' traffic (12 of 15). The OpenAPI file documents a 202 pending result with an `NVCF-REQID` to poll, but no 429, Retry-After or backoff guidance, and NVIDIA's FAQ says to expect long waits under load (3 of 15). No SLA for the trial. NVIDIA's SLAs come with AI Enterprise for self-hosted software (0). The card says v1.0 GA, but the hosted page calls the model early access and NVIDIA calls the catalogue a trial for prototyping only (3 of 10). 58.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 3. Payments & pricing, 35 out of 100, up to 8.1 more on the total

Why it scored 35: No x402, MPP or L402 (0). The hosted endpoint is free and says so without a login, and the production routes are priced in public (AI Enterprise at $4,500 a GPU a year, partner hosts per million tokens), but the free endpoint can't be used in production and has no per-call price of its own (15 of 20). Free with no card per build.nvidia.com/llms.txt (20). A person signs in at build.nvidia.com, which uses SMS verification per the status page, to generate a key. No keyless or programmatic route (0). 35.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 4. Agent ergonomics, 65 out of 100, up to 5.7 more on the total

Why it scored 65: Model reading of the checklist (tool use, structured output, caching, context, batch, SDKs, errors). OpenAI-format tools, `tool_choice` auto, none, required or a named function, and parallel calls on by default (20 of 20). `response_format` with `json_schema` and `strict` in the OpenAPI file. Not tested by us (12 of 15). The card describes prefix caching for self-hosted serving, but no caching on the hosted endpoint is documented (2 of 10). 1M tokens in the weights, but the hosted page lists 262K, and completions are capped at 131,072 tokens (12 of 15). No batch mode (0 of 10). Any OpenAI SDK works with the base URL, and NVIDIA maintains `langchain-nvidia-ai-endpoints`. Only `messages` is required, but reasoning defaults to on at high effort and the reference defaults `stream` to true (10 of 15). Errors carry a message, type and request id, with no 429 or rate-limit body documented (9 of 15). 65.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Schema & documentation, 74 out of 100, up to 4.2 more on the total

Why it scored 74: Model reading. An OpenAPI 3.1 file for this model's `/chat/completions` on docs.api.nvidia.com, with typed request and response schemas (25). build.nvidia.com/llms.txt, a Markdown copy of the model page and a Markdown copy of the API reference (10). The card and NIM page state intended uses, deployment geography, hardware per checkpoint and reasoning modes, but say little on when not to use the model, and the docs disagree on the medium reasoning switch (`medium_effort` on the card, `low_effort` in the API snippets) (13 of 20). Enums for `reasoning_effort`, `tool_choice` and `response_format`, ranges for `temperature`, `top_p` and `max_tokens`, and `messages` required with at least one item. `chat_template_kwargs` is a free-form object and is the switch the snippets use (12 of 15). Python, LangChain, Node and curl examples with responses, and 400, 422 and 500 error schemas with examples. No 429 (11 of 15). No changelog for the catalogue or the model. The OpenAPI version is 1.0.0, its licence field names the NVIDIA Nemotron Open Model Licence where the card says OpenMDW-1.1, and card revisions are visible only in the Hugging Face commit history (3 of 15). 74.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Transparency & trust, 59 out of 100, up to 3.6 more on the total

Made of editorial 38, provenance 79.

Why it scored 59: Weights under OpenMDW-1.1, a short permissive licence from the Linux Foundation with no field-of-use limits, but not an OSI licence. Datasets and training recipes are released. The hosted trial is under terms we couldn't read, and the OpenAPI file names a different model licence and the AI Enterprise EULA (22 of 30). NVIDIA's privacy policy (22 September 2025) doesn't address API prompts. The model page says use is logged for security and to improve NVIDIA products, with improvement data not linked to identity, and gives no retention period. The trial terms that may say more were unread (8 of 30). No deprecation policy or dated notices for catalogue models. The page carries an `isDeprecated` flag only (3 of 20). The quickstart says hosted endpoints run on NVIDIA DGX Cloud, but no subprocessors or data locations for the catalogue were found (5 of 20). 38.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Privacy policy: read, states 7 of the 8 things a reader expects (9.3 of 10)
- Changelog: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 71 out of 100, up to 2.5 more on the total

Why it scored 71: Read for a model released two days ago. Released on 15 October 2026, with card revisions to 16 October (30). NVIDIA released three Nemotron 3.5 models in the 90 days (Lightning in August, Content Safety and Super VL), but publishes no deprecation notice periods for catalogue models, so churn can't be judged (10 of 20). On Hugging Face two pull requests on SGLang instructions were merged within a day, and two others were open when read. Support for the free tier is the developer forum (15 of 25). NVIDIA maintains `langchain-nvidia-ai-endpoints`, and the endpoint takes OpenAI SDKs (10 of 15). Day-one recipes for vLLM 0.32.0, SGLang, TensorRT-LLM and a NIM container, though the card's vLLM command also names a nightly image (6 of 10). 71.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: the NVIDIA API Trial Terms of Service PDF at assets.ngc.nvidia.com, whose robots.txt answered with an S3 AccessDenied error, so data use, retention and acceptable-use terms for the hosted trial are unread
- Whether the consent text on the model page (use logged for security and to improve NVIDIA products) covers API calls or only the in-browser playground. It sits in the page's playground copy
- Hosted context differs between pages. build.nvidia.com lists 262K, OpenRouter's NVIDIA-served free route lists 1,000,000, and the card says up to 1M. Not tested
- Whether the hosted endpoint accepts `reasoning_effort` directly. The API reference says NVIDIA's snippets translate it into `chat_template_kwargs`, so the agent notes use `chat_template_kwargs`
- The OpenAPI file's licence field names the NVIDIA Nemotron Open Model Licence and its terms link the AI Enterprise EULA, while the card and model page name OpenMDW-1.1 and the API Trial Terms. Recorded as found
- The launch post's 'highest accuracy' for always-on agents and '5x faster than the next best' are NVIDIA's claims. Its own table shows Qwen3.8-Flash-Next and GLM-5.3-Flash ahead on most benchmarks shown, but the post limits the claims to its class and to specialised agents, so no deduction is taken
- Bolt 74B-A9B 24GB is covered in details, as the scout proposed. It has its own Hugging Face repository, a different context limit (256K) and no hosted endpoint, and is aimed at local agents, so it could take its own local-ai listing if the founder wants one. Bolt E4B is due on 23 October 2026
- The scout proposed the slug `nvidia-nemotron` for a family listing. This listing is `nemotron-3-5-super-vl` as the batch list proposed. Nemotron 3.5 Lightning, Nemotron 3 Ultra, Nemotron 3 Super and Nemotron 3 Nano Omni are on build.nvidia.com and not listed. NVIDIA-Nemotron-Labs-Dory1.0-4B was not covered
- Rate-limit numbers come from build.nvidia.com's page copy. Forum users report the same 40 requests a minute, which we didn't rely on
- GitHub was not read (the NeMo cookbook and Gym recipes the card links), so no repository CI or issues were checked

## Weaknesses

- NVIDIA says its hosted endpoints are for prototyping, research and testing only, and production use needs a partner endpoint or self-hosting
- Rate limit of up to 40 requests a minute and 10,000 a day, which NVIDIA says can vary with other users' traffic
- The NVIDIA API Trial Terms of Service, which govern the hosted endpoint, are on a host whose robots.txt answered access denied, so we didn't read them
- The hosted page lists 262K tokens of context while the card says 1M, and labels the model early access while the card says GA
- No changelog, deprecation policy or SLA for the hosted catalogue was found

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Set `chat_template_kwargs` to `{"enable_thinking": false}` for short tasks. Reasoning is on by default, and reasoning tokens share the `max_tokens` budget
- Read reasoning from `reasoning_content` and the answer from `content`, and set `stream` explicitly since the API reference defaults it to true
- Don't send confidential or personal data to the hosted trial. The model page asks callers not to and says use is logged
- 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
- 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

## What the review panel asked for

- a notice period for catalogue models
- Publish a first-party production price
- Document the 429 response
- Reconcile context and release status
- Document 429 and Retry-After behaviour
- Publish a changelog for catalogue models

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.