Nemotron

Two hybrid MoE text models from NVIDIA's Nemotron 3 line.

Nemotron in this matrix is a pair of hybrid text models: Lightning at 30B total / 3B active, and Super at 120B total / 12B active. Both interleave Mamba-2, MoE, and attention, and both are validated on the same Strix Halo snapshot as the dense and vision-language entries.

Nemotron 3.5 Lightning 30B-A3B

  • Kind: Language model
  • Task: Text generation, agentic
  • Parameters: 30B total, 3B active
  • Architecture: Hybrid Mamba-2 + sparse mixture of experts
  • Quantization: int4 RTN
  • Base model: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
  • Validation status: New in v0.5.3
  • Measured in v0.5.3: 73.06 tokens/s, 1.83 s to first token, at a 2048-token prompt — all three lengths

Nemotron 3 Super 120B-A12B

  • Kind: Language model
  • Task: Text generation, agentic
  • Parameters: 120B total, 12B active
  • Architecture: Hybrid LatentMoE (Mamba-2 + MoE + attention)
  • Quantization: int4 RTN, quantized MoE
  • Base model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
  • Validation status: New in v0.5.3
  • Measured in v0.5.3: 22.19 tokens/s, 10.82 s to first token, at a 2048-token prompt — all three lengths