NVIDIA-Nemotron-Labs-Teacher-General-Reasoning: The 550B Reasoning Teacher Behind Nemotron 3 Ultra Just Went Open Weights

On August 14, 2026, NVIDIA published NVIDIA-Nemotron-Labs-Teacher-General-Reasoning on Hugging Face — a 550B-parameter reasoning model that, until yesterday, existed only inside the training pipeline of its flagship Nemotron 3 Ultra. It is the "general reasoning" teacher from the Multi-Teacher On-Policy Distillation (MOPD) recipe NVIDIA used to train the 550B-A55B Ultra student, and the release matters for a simple reason: in the Nemotron 3 Ultra technical report NVIDIA shipped in June, the training recipe stated plainly that the per-teacher checkpoints would not be open-sourced. This one now is — weights, tokenizer, chat template, and serving configs included, under the commercial-friendly OpenMDW-1.1 license. A sibling, NVIDIA-Nemotron-Labs-Teacher-STEM, dropped the same day, which reads less like a one-off and more like the teacher panel starting to come out.

What a teacher model actually is

MOPD is the post-training technique that gives Nemotron 3 Ultra its agentic reasoning. Instead of distilling a single large teacher into a student, NVIDIA trained more than ten domain-specialized teachers — for reasoning, research, legal analysis, software engineering, tool use, and other areas — then let the student generate its own rollouts and used each teacher to score those rollouts in its area of expertise. Because scoring happens on the student's own on-policy outputs rather than on a fixed offline dataset, the training signal stays aligned with what the student actually produces at inference time. NVIDIA calls the loop co-evolution: new teacher rounds are initialized from updated student checkpoints, so both sides keep improving.

This checkpoint is the general-reasoning member of that panel. It is produced from the post-trained Nemotron 3 Ultra student through an extra round of reasoning-intensive SFT and reinforcement learning, and the model card describes it as optimized for extended, high-quality reasoning traces on the hardest multi-step problems across math, logic, and abstract reasoning — including formal proofs and competitive coding. Inside MOPD it plays several roles at once, one checkpoint, many roles: reasoning-trace generation, math judging, and the equivalence judge that grades free-form short answers against a reference. NVIDIA also ships it standalone because it is a strong reasoner in its own right — intended for long-horizon reasoning-trace generation, hard problem solving, and acting as a teacher or grader inside your own distillation pipeline.

The specs in one pass

• Parameters — 550B total / 55B active, sparse LatentMoE

• Architecture — Mamba2-Transformer hybrid latent mixture-of-experts with Multi-Token Prediction (MTP)

• Context window — up to 1M tokens; 256K default in the reference serving configs

• Languages — 10: English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, Chinese

• License — OpenMDW-1.1, commercial and non-commercial use allowed

• Minimum hardware — 4×B200 (NVFP4 weights); GB200/GB300 and H100-class also supported

The architecture is the same family as Nemotron 3 Ultra itself: the 550B-A55B shape with interleaved Mamba-2 and MoE layers, select attention layers, and MTP heads for native speculative decoding. The teacher is not a smaller student — it is a differently-trained variant of the same stack, specialized for long-horizon reasoning rather than general agentic work.

Why open-sourcing a teacher matters

Teacher checkpoints are the rarest kind of open model release. Companies publish students — the models you deploy — and datasets all the time; they almost never publish the models that graded the students' work. That is why the June report's line that per-teacher checkpoints would not be released was read as closing the door on reproducing the MOPD pipeline end to end. This release cracks that door open, at least for the reasoning teacher.

For teams building their own reasoning models, that is concrete value. The reward signal that shaped Nemotron 3 Ultra's reasoning was partly written by this checkpoint, and now it can be studied directly, run as a judge, or used to generate long-horizon reasoning traces for SFT data. Combined with the already-open post-training data (nvidia/nemotron-post-training-v3) and the published training recipes, it narrows the gap between "NVIDIA trained a great model" and "we can train one like it."

The thinking dial

One distinctive feature carries over from the Nemotron 3 family: reasoning is a switch, not a default. The chat template accepts enable_thinking=true/false, a medium_effort option, and a reasoning_budget token ceiling, so a caller can force deep long-horizon reasoning when a problem needs it and get direct answers — faster and cheaper — when it does not. For a teacher model that matters more than it looks: distillation runs want cheap grading passes on easy samples and expensive reasoning traces on hard ones, and a single checkpoint that supports both modes removes the need to swap models mid-pipeline.

Running it

This is not a model you will call casually. The minimum sensible serving setup is 4×B200 with NVFP4 weights and an fp8 KV cache; the reference configs also cover multi-node GB200/GB300 and H100-class boxes. NVIDIA publishes cooking guides for three engines — vLLM (0.22), SGLang (0.5.13), and TensorRT-LLM (1.3.0rc17) — all exposing an OpenAI-compatible API on port 8000, with MTP or EAGLE speculative decoding and a flashinfer Mamba backend. The 1M-token context requires an explicit allow-flag in each engine (VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and equivalents); the default ceiling is 256K.

As of writing, the model is not deployed by any inference provider on Hugging Face and NVIDIA has not announced NIM hosting — this is a weights-only release. That makes it a model for labs that already run large GPU fleets, or for teams whose distillation pipeline specifically needs the exact teacher signal. When it does land on a managed API, the pricing math is simple: this is a 55B-active MoE, and whoever hosts it charges what the GPUs cost. On a routing layer that works at 0% markup, you pay the provider's list price with no platform fee on top — and the same key that reaches this model reaches the frontier models you compare its traces against, with automatic failover if a host drops. That is the part a router like OrcaRouter is for; the teacher itself is the part you self-host.

The honest caveats

This is a 24-hour-old checkpoint, and the model card contains zero benchmark numbers. That is not an omission in this write-up — it is the state of the release. NVIDIA published architecture and training details but no evaluation table, and no independent lab has had time to run it yet. The nearest reference point is the student's vendor-reported numbers: Nemotron 3 Ultra reports 71.9 on SWE-Bench Verified, 86.8 on MMLU-Pro, 89.0 on LiveCodeBench v6, and roughly 95 on RULER at 1M context. Those describe the Ultra student, not this teacher, and they are NVIDIA's own figures. Anyone planning a distillation run on this checkpoint should treat its reasoning quality as unverified until independent scores appear.

Two more caveats are baked into the card. The post-training corpus is heavily English-weighted — about 8.6 million English samples against roughly 138,000 for each of the other nine languages — so multilingual reasoning quality should not be assumed from the 10-language tag. And the card itself calls out representation skew in the base training data and recommends bias audits before downstream use.

Who should care

Three groups get real value here. Distillation researchers finally have an actual MOPD teacher to dissect, not just a recipe describing one. Labs training their own reasoning models get a long-horizon trace generator and a judge that shipped from the same post-training lineage as the model they are trying to match. And anyone running on-policy RL can use it the way NVIDIA did — as a scorer for the hardest problems, with thinking switched on only where it pays for itself.

If you are in none of those groups, this release changes little for you today: the model is big, unproven, and self-host-only, and the fastest way to act on the underlying trend — more open reasoning teachers appearing — is to keep your pipeline's model layer swappable behind one interface rather than welded to any single checkpoint.

What to watch next

Three things will tell whether this is a one-off or the panel coming out. First, whether sibling teachers follow the same path — NVIDIA-Nemotron-Labs-Teacher-STEM already landed the same day, so the question is which domain specialists come next and whether they are weighted the same way. Second, whether NVIDIA NIM or a managed host starts serving them, which would turn a labs-only release into something the rest of the industry can actually call. Third, whether independent benchmarks show up — an Artificial Analysis entry or LMArena run would be the first real check on whether the teacher's reasoning quality matches the student it trained.