NVIDIA Publishes the Nemotron Recipe Behind Gold-Level IMO and IOI Results

NVIDIA connected two Nemotron competition systems on October 7

NVIDIA’s October 7 Hugging Face publication puts its 2026 mathematics and competitive-programming work under one claim: Nemotron 3 can be specialized with supervised fine-tuning, reinforcement learning where useful, and feedback-driven inference. It reports 535.4 out of 600 for an IOI system and 30 out of 42 for an IMO system, both above the cited gold thresholds of 361.12 and 29.

The availability date needs a qualifier. The linked papers and specialist repositories appeared in September; October 7 is the publication date of the unified account, not the first release date of a new base model. NVIDIA names no new managed API SKU or price in this announcement. What readers can access now is a set of public weights, datasets, code and papers under their respective terms.

The IOI result came from Ultra-CC plus GenCorrect

The coding checkpoint is nvidia/NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4: a 550-billion-total, 55-billion-active specialist based on Nemotron 3 Ultra, with a stated context limit of 262,144 tokens. NVIDIA says it fine-tuned the model for one epoch on 477,642 synthetic reasoning traces distilled from GLM-5.2 across 22,000 competitive-programming problems, then quantized it to NVFP4.

The reported 535.4/600 is not a single-response checkpoint score. NVIDIA paired Ultra-CC with GenCorrect, which generates candidate programs, clusters diverse candidates, obtains evaluator feedback and refines later rounds within a submission budget. The paper describes one prospective run under the competition’s time, internet and submission constraints before public problem release. It was unofficial, unsupervised and not included in the IOI ranking; AccessAllGPT did not independently verify the setup or result.

The IMO result used three checkpoints and a large search

For mathematics, NVIDIA released Nemotron-3-Labs-Ultra-Math-SFT and Nemotron-3-Labs-Ultra-Math-RL alongside the general Nemotron 3 Ultra checkpoint. The SFT corpus contains 414,890 quality-filtered examples across 15,818 unique proof problems; the RL dataset contains 9,597 problems. NVIDIA says the final system generated, verified and refined natural-language proofs without a formal prover, external tools or internet access.

The released recipe makes the compute shape concrete. Its first round schedules eight prompts across three checkpoints with 16 samples each—384 proof attempts per problem. Every distinct proof can receive 16 verifier judgments; refinement can continue through round eight; and each finalist can receive 48 grading judgments. Official IMO graders awarded the submitted set 30/42, according to the authors. That supports a system result, not a claim that either specialist alone scores 30/42 or that a low-budget deployment will match it.

The open release is substantial but not turnkey

The IMO collection exposes two specialist checkpoints, SFT and RL datasets, Nemotron-IMO-Bench with 200 olympiad-level problems, and the paper. NeMo Skills supplies prompts, submitted proofs, configuration and an auditable pipeline with request identities, hashes, manifests, errors and resume behavior. The coding model card supplies weights, a vLLM deployment recipe and its reported training and benchmark details.

All three specialist repositories were public and ungated when reviewed. Their cards use the OpenMDW 1.1 license, so “open weights” should not be collapsed into “Apache licensed” or “unrestricted.” Hugging Face metadata showed roughly 352 GB stored for the NVFP4 coding repository and roughly 1.12 TB for each BF16 mathematics repository. The mathematics card recommends eight B200 GPUs for a single-node deployment; buyers still need infrastructure, serving, energy and inference-search cost estimates.

There is no single public price or hosted-access promise

NVIDIA’s release does not quote a managed price for either specialist system. The mathematics model records listed no hosted inference provider in the Hugging Face API at retrieval. The open recipe can target a hosted OpenAI-compatible endpoint or local vLLM servers, but an example endpoint in configuration documentation is not proof that every named specialist is currently offered there or covered by one commercial rate.

Procurement should therefore separate artifact access from service access. For self-hosting, calculate storage, GPU topology, model loading, long-context memory, candidate count, verifier count, retries and final judging. For a hosted service, require the exact model IDs, context and output limits, throughput, retention, region, rate limits and token pricing in current service documentation before estimating cost.

The results show specialization leverage, not broad superiority

Across both projects, the useful pattern is co-design. Domain data changed the checkpoint; specialist and general checkpoints supplied different candidates or judgments; and test-time search converted those capabilities into stronger final submissions. This is more informative than attributing the outcome to parameter count or fine-tuning alone.

Transfer remains unproven. Competitive-programming tasks have executable hidden tests; production software work includes changing repositories, requirements, tools, security boundaries and maintenance. Olympiad proofs are difficult but bounded; research mathematics and formal verification have different evidence requirements. Neither result establishes chat quality, agent reliability, factual accuracy, safety, latency or cost effectiveness on a buyer’s workload.

What builders should do next

Start by choosing the relevant artifact, not the medal headline. A coding team can compare the released Ultra-CC checkpoint with its existing model on frozen, uncontaminated tasks and then add GenCorrect one round at a time. A math team can first reproduce the published pipeline’s dry run, inspect the prompts and submitted proofs, and evaluate the SFT and RL checkpoints separately before attempting the full ensemble.

Adopt the artifacts for research when the license, hardware and bounded task fit; constrain any pilot to offline workloads with executable or expert review; wait if hosted access, cost or reproducibility is unresolved; and reject using the two medal labels as a general production-model ranking. Record both quality and total attempts, tokens, wall time, hardware, failures and human review so the decision reflects the system that produced the answer.

Disclosures

AccessAllGPT received no NVIDIA or Hugging Face access, weights, GPU time, hosted credits, dataset preview, briefing, source code, competition submission, review, payment or compensation for this article. NVIDIA and Hugging Face did not sponsor, review or endorse it. AccessAllGPT did not run, score or rank the systems. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with NVIDIA, Hugging Face, arXiv, IMO or IOI. Publication-wide relationships are listed on the disclosures page.