Not established from the available sources.
NVIDIA · MODEL RELEASE TRACKER
Nemotron 3
Nemotron 3 is an NVIDIA hybrid model that interleaves state-space model (SSM) layers with a few attention layers. Hybrid SSM implementations mentioned for such models include Mamba-2, Gated DeltaNet (GDN), or Kimi Delta Attention (KDA).CURRENT SNAPSHOT1/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Not established from the available sources.
IOI evaluation was a live, prospective run under human-like contest constraints and was unofficialThe IOI result came from a live, prospective run under the same time, internet-access, and submission constraints as human contestants; it was an unofficial, unsupervised benchmark and was not included in the official IOI ranking.
IMO system proofs were graded by official IMO gradersThe IMO system’s submitted proofs were graded by official IMO graders.
IOI score progression for Nemotron-3-Nano-CC and Nemotron-3-Ultra-CCNano improved from 130 points before post-training to 280 after SFT and 291 after RL; with GenCorrect it reached 468 points (crossing the gold threshold of 438.3). Ultra-CC reached 502 points using the same test-time strategy.
Not established from the available sources.
VERIFIABLE FACTS
Every value stays attached to a source and date.
CAPABILITIES · Fine-tuned specialists from Nemotron 3 reached gold-medal level at IOI 2026 and IMO 2026DEVELOPER CLAIM
Starting from Nemotron 3, teams used supervised fine-tuning, reinforcement learning, and feedback-driven inference to create systems that reached gold-medal level at both IMO 2026 and IOI 2026.
BENCHMARKS · IOI evaluation was a live, prospective run under human-like contest constraints and was unofficialDEVELOPER CLAIM
The IOI result came from a live, prospective run under the same time, internet-access, and submission constraints as human contestants; it was an unofficial, unsupervised benchmark and was not included in the official IOI ranking.
BENCHMARKS · IMO system proofs were graded by official IMO gradersDEVELOPER CLAIM
The IMO system’s submitted proofs were graded by official IMO graders.
CAPABILITIES · Named Nemotron-3 variants, parameter counts, and training appliedDEVELOPER CLAIM
For competitive programming, teams trained two specialists: Nemotron-3-Nano-CC (30 billion total parameters, 3 billion active parameters) which received SFT and RL, and Nemotron-3-Ultra-CC (550 billion total parameters, 55 billion active parameters) which received SFT.
BENCHMARKS · IOI score progression for Nemotron-3-Nano-CC and Nemotron-3-Ultra-CCDEVELOPER CLAIM
Nano improved from 130 points before post-training to 280 after SFT and 291 after RL; with GenCorrect it reached 468 points (crossing the gold threshold of 438.3). Ultra-CC reached 502 points using the same test-time strategy.
CAPABILITIES · Architecture (hybrid SSM + attention)DEVELOPER CLAIM
Nemotron 3 is built as a hybrid model that interleaves state-space model (SSM) layers with a few attention layers.
CAPABILITIES · SSM layer implementationsDEVELOPER CLAIM
Nemotron 3 (as a hybrid model) uses SSM layers; examples mentioned for such models include Mamba-2, Gated DeltaNet (GDN), or Kimi Delta Attention (KDA).
LIMITATIONS · SSM-layer latency behavior / limitationDEVELOPER CLAIM
Because Nemotron 3 uses SSM layers, at serving batch sizes the SSM layers (rather than attention) can become the latency bottleneck for up to about 100,000 tokens of context.
WHAT CHANGED
Stored passport versions, without reconstructed history.
CAPABILITIES · Architecture (hybrid SSM + attention)+ Nemotron 3 is built as a hybrid model that interleaves state-space model (SSM) layers with a few attention layers.
LIMITATIONS · SSM-layer latency behavior / limitation+ Because Nemotron 3 uses SSM layers, at serving batch sizes the SSM layers (rather than attention) can become the latency bottleneck for up to about 100,000 tokens of context.
CAPABILITIES · SSM layer implementations+ Nemotron 3 (as a hybrid model) uses SSM layers; examples mentioned for such models include Mamba-2, Gated DeltaNet (GDN), or Kimi Delta Attention (KDA).