NVIDIA-Nemotron-Labs-Teacher-Chat: 550 billion parameters to train a smaller model
7 min read
On 14 August 2026, within about ninety seconds, NVIDIA published five repositories on Hugging Face: NVIDIA-Nemotron-Labs-Teacher-STEM, -Instruction-Following, -General-Reasoning, -Competition-Coding and, last, -Chat. We verified this ourselves on 17 August, querying the public Hugging Face API. Repository examined: nvidia/NVIDIA-Nemotron-Labs-Teacher-Chat, createdAt 2026-08-14T17:09:11Z, last modified the same day at 21:24. At that moment: 1,141 downloads, 0 likes, 238 files, of which 225 .safetensors shards. Declared tags: nemotron_h, text-generation, nvidia, pytorch, nemotron-3, latent-moe, mtp, conversational, custom_code, plus twelve languages and two datasets. The gated field is false: we tested it by downloading config.json without authenticating — response HTTP 200, no login required.
A model not built for you, but for the model it trains
The card states it in the first line of its summary table: 550 billion total parameters, 55 billion active, LatentMoE architecture — a hybrid of Mamba-2 layers, Mixture-of-Experts and a handful of attention layers, with Multi-Token Prediction. Declared context length up to 1 million tokens (256,000 by default in the card’s example configurations). Minimum hardware: 4×GB200, 4×B200, 4×GB300, 4×B300, or 8×H100 — for a single inference session, not for training from scratch.
It is not built to answer your questions, even though it can: it is a teacher, one of “more than ten domain-specialized teachers that supply training signal to Multi-Teacher On-Policy Distillation (MOPD)” — the stage NVIDIA uses to produce the family’s flagship model, Nemotron 3 Ultra. Of those internal teachers, five were released publicly as standalone checkpoints on 14 August — “released as a standalone checkpoint because it is a strong chat model in its own right,” the card says of the Chat variant. In practice: NVIDIA is releasing not only the model you would use, but the tool for generating the data to train your own.
For anyone who wants a smaller model specialised to their own domain, without sending prompts and documents to an external API, a downloadable teacher is the difference between two worlds: generating training traces inside your own perimeter, or outside it. We wrote about this before regarding unauthorised distillation from commercial API outputs. Here the problem is reversed: the licence explicitly permits it, and the constraint is entirely technical — starting with eight H100 GPUs just to run the teacher itself.
The licence, line by line
license: other, license_name: openmdw-1.1, linking to https://openmdw.ai/license/1-1/. This is not text NVIDIA wrote for the occasion: it is OpenMDW-1.1, the licence the Linux Foundation published on 28 May 2026, which we had already read in full for another model, poolside’s Laguna S 2.1. We find it again here because the card cites it again, verbatim.
The grant is broad but conditional, and the condition sits in its opening words: “Subject to your compliance with this agreement, permission is hereby granted, free of charge, to deal in the Model Materials without restriction, including under all copyright, patent, database, and trade secret rights included or embodied therein”. No revenue threshold, no prohibited field of use, no excluded geography.
Two obligations, both procedural. Whoever redistributes retains “(1) a copy of this agreement, and (2) all copyright notices and other notices of origin included in the Model Materials that are applicable to your distribution”. And whoever sues loses everything: if you bring a claim alleging the Model Materials infringe a patent or copyright, “all rights and grants made to you hereunder are terminated, unless that lawsuit was in response to a corresponding lawsuit first brought against you”.
The line legal counsel should read first is another one: “YOU ARE SOLELY RESPONSIBLE FOR (1) CLEARING RIGHTS OF OTHER PERSONS THAT MAY APPLY TO THE MODEL MATERIALS.” No supplier indemnifies you. And on generated outputs: “This agreement does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs” — including, here, the synthetic data this model generates to train another one.
The subcards the model card promises and does not deliver
The card closes its ethics section by pointing to four “Subcards”: Safety, Explainability, Bias and Privacy, each linked as an .md file inside the repository. We opened each one on 17 August: safety.md, explainability.md, bias.md and privacy.md all return HTTP 404. They do not exist — probably a card fragment copied from another model in the same family and never adapted: even the accuracy_plot.png image, cited just above, gives the same error. It is the most honest possible answer to the question of a second document: there is no separate use-restriction file beyond the licence, but there is also no safety documentation the card itself claims to have. A technical report ought to say more than this one does.
See the service · Talk to an engineer
A register cannot stop at the repository’s name
For a register of models in production, “NVIDIA-Nemotron-Labs-Teacher-Chat, openmdw-1.1” is not a complete line. A gated repository opened with a single click — the subject of another note of ours, published this same day — shows that the distribution channel matters as much as the licence text. Here the channel is open without an account, verified with a real download. What still needs recording: the exact checkpoint version (v1.0 - Ultra Chat Teacher GA, the card says), the digest of the shards downloaded, the acquisition date, the list of promised and missing subcards, and, given its role as a teacher, the intended use — generating data to train an internal model, not serving end-user requests. These are different boxes from the ones a model that simply answers a prompt requires.
What we do not know
We did not download the 225 shards or run the model: we make no claim about the quality of its answers, nor about the quality of the data it would generate for a model distilled under its guidance. Benchmarks appearing elsewhere in the Nemotron 3 family are not ours, and we do not report them as verified. NVIDIA also offers hosted access on build.nvidia.com, which responds with HTTP 200: we did not query it with an account, so we do not know whether it offers functionality absent from the downloadable weights. We do not know what hardware a production use would really require beyond the stated minimum, nor the cost of a real distillation session. We offer no legal qualifications: whether generating training data with this model falls under a given AI Act category depends on the concrete use made of it, and that assessment belongs to whoever puts it into service.
How we solve this
Compliance: the register of models in production becomes a control running on the client’s own systems — for a teacher like this one, the licence line is not enough: you need the checkpoint version, the shard digests, the acquisition channel and date, the list of what the card promises and fails to deliver, and the approved use — generating data to train an internal model — together with who signed off on it in writing. That is the trail to produce at an inspection, to a client during a tender, or to a board.
Decisions: the same set-up brings together models, data, contracts, archives and documents into a single operating model on which AI agents execute decisions with a human operator in command, for large enterprises, defence, government and healthcare. A teacher with 550 billion parameters, eight H100 GPUs as its floor, is direct proof that downloadable weights without the compute that makes them runnable inside your own perimeter are not sovereignty: the licence is free, the compute to exercise it is not. Always in both delivery modes — on-premise, on self-contained machines that do not require deep integration into the client’s network, or a dedicated cloud with a dedicated VPN and a data centre in Italy — and always with shared governance: you do not need to already have, in-house, someone who administers a set-up of this scale.
From the first session, at no cost, comes the dated register of the models in service: for each one, how it was obtained, whether the channel is still open, and what you have kept — including the boxes that remain empty. It stays yours even if we do not go on to work together.