Model reference · open weights
NVIDIA-Nemotron-Labs-3-Competitive-Coding is an open-weight language model from NVIDIA. NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 (NVFP4) weighs 352 GB; it needs more than the reference configurations below.
What it is
| Released by | NVIDIA |
|---|---|
| Type | Language models |
| Task | Text gen · MoE |
| Parameters (lead) | 335.0B |
| Runs with | transformers |
| Released | 2026-09-04 |
| Popularity | 0 downloads / month |
| Weights | 352 GB (NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 (NVFP4), file size) |
| Licence | Its own licence terms |
What it runs on
Weights 352 GB (file size) · runtime overhead from 762 MB on a small card.
How much memory each request adds is not estimated yet for this architecture — only the weights are. They need the cards below at the least, plus room for the context.
| Card | The weights alone |
|---|---|
| RTX 3060 12 GB … 8× B200 180 GB 16 smaller cards | does not fit |
From the model card
| Total Parameters | 550B (55B active) |
| Architecture | Based on Nemotron-3-Ultra |
| Context Length | Up to 262,144 tokens |
| Best For | Competitive programming, algorithmic problem solving, code-reasoning research and benchmarking, test-time-compute research |
| Reasoning Mode | Structured Explanation → Confidence → Answer response format with step-by-step reasoning before the final C++ solution |
| License | OpenMDW License Agreement, version 1.1 |
| Release Date | Hugging Face: 09/03/2026 via model page |
This checkpoint is a competitive-programming specialist model, not a general-purpose chat or agent model. It supports commercial and non-commercial applications, including research and evaluation contexts such as competitive programming benchmarks, code-reasoning research, and test-time-compute studies (for example, GenCorrect-style iterative refinement). See Use Case below.
Model Developer: NVIDIA Corporation
Model Development: Fine-tuned from NVIDIA-Nemotron-3-Ultra-550B-A55B
NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.
Nemotron-Labs-3-Competitive-Coding is a competitive-programming specialist model based on Nemotron-3-Ultra, fine-tuned for one epoch on 477,642 synthetic reasoning traces distilled from GLM-5.2 across 22,000 curated problems spanning 16 regional and international competitive-programming contest families. Selected as the SFT teacher for its higher accuracy and roughly 30% shorter generations compared to a DeepSeek-V4-Flash-trained variant, GLM-5.2 distillation yields a model that, combined at inference time with GenCorrect — an iterative closed-loop test-time compute strategy that generates diverse candidate solutions, incorporates evaluator feedback, and refines subsequent generations under a fixed submission budget — was evaluated live and prospectively on the IOI 2026 problem set under official contest time, internet-access, and submission constraints, scoring 535.4 out of 600 and surpassing both the gold-medal threshold (361.12) and the top human contestant's score (498.27), making it the first AI system reported to outscore the highest-scoring human contestant on an IOI problem set.
This model is ready for commercial or non-commercial use.
Governing Download Terms: Use of this model is governed by the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1).
| Benchmark | Nemotron-Labs-3-Competitive-Coding |
|---|---|
| IOI 2025 — with GenCorrect (5 rounds) | 502.0 / 600 |
| ICPC 2025 — with GenCorrect (5 rounds) | 9.6 / 12 problems solved |
| LiveCodeBench Pro — Pass@1 | 74.5% |
| IOI 2026 — live, prospective, competition-specific run | 535.4 / 600 (Gold; exceeds gold threshold of 361.12 and top human score of 498.27) |
All results are from the source paper, Post-Training Language Models for Gold-Medal Performance in Coding Competitions (NVIDIA, arXiv:2609.02849). IOI Score@1/Score@200 and GenCorrect results are averaged over multiple independent runs; see the paper for full methodology. IOI 2025 was used as a development benchmark; IOI 2026 results are from a single prospective live run conducted under official IOI time, internet-access, and submission constraints before problems were publicly released, and were not part of the official IOI rankings.
Nemotron-Labs-3-Competitive-Coding is intended for researchers and developers evaluating or advancing frontier code-reasoning capability, particularly on competitive programming and algorithmic problem solving where a solution must satisfy strict correctness, efficiency, and hidden test-case constraints. It is suited to benchmarking and research on long-horizon reasoning, agentic code generation, and test-time compute strategies such as GenCorrect-style iterative refinement, rather than general-purpose chat, instruction-following, or production coding-assistant deployment. Use in safety-critical or real-time production systems requires further evaluation and safeguards appropriate to the application.
Hugging Face: 09/03/2026 via model page
This model inherits its architecture unchanged from Nemotron-3-Ultra. See the Nemotron-3-Ultra Technical Report for architecture details.
Input Type(s): Text
Input Format(s): String
Input Parameters: One-Dimensional (1D)
Other Properties Related to Input: Maximum context length up to 262,144 tokens
Quoted from the model card on Hugging Face — the full card is behind the Hugging Face link above.
How it works