Nemotron-Cascade-2-30B-A3B-FP8

FP8 W8A8 quantization of Nemotron-Cascade-2-30B-A3B — a 32B hybrid Mamba-MoE reasoning model (3.2B active). Quantized with NVIDIA ModelOpt using the same selective quantization recipe as NVIDIA's official Nano FP8: MoE experts and Mamba GEMMs in FP8, attention and sensitive layers in BF16, KV cache in FP8.

Benchmarks

Calculated using NVIDIA-NeMo/Evaluator with config from Nemotron-3-Super-120B's eval config:

Benchmark	Nemotron-Cascade-2-30B-A3B (reproduced results)	Nemotron-Cascade-2-30B-A3B-FP8 (this model)
AIME 2025 (avg@8)	98.8	96.7
AIME 2026 (avg@8)	94.2	95.0
HMMT Feb 2025 (avg@8)	92.9	93.8

With the low sample count (8 rollouts per problem), a deviation of ±2% accross runs is expected. Safe to say BF16 and FP8 variants are equivalent in reasoning performance.

Quantization Details

• Method: FP8 Post-Training Quantization (PTQ), without Quantization-Aware Distillation (QAD)
• Format: FP8 E4M3 — static per-tensor weight scaling, dynamic per-token activation scaling at inference
• KV cache: FP8
• Tooling: NVIDIA ModelOpt

Selective Quantization Recipe

Follows the Nano-architecture selective quantization recipe from the Nemotron 3 Nano Technical Report (Section 4). Same recipe as NVIDIA's official FP8 checkpoint. Sensitive components are kept in higher precision:

Component	Precision	Rationale
MoE expert GEMMs (routed + shared)	FP8	All 23 MoE layers, 128 routed + 2 shared experts each
Mamba GEMMs (non-adjacent)	FP8	17 of 23 Mamba layers
Attention layers (all 6)	BF16	Most sensitive — kept BF16 per NVIDIA sensitivity analysis
Mamba layers adjacent to attention (6)	BF16	Layers {4, 11, 18, 25, 32, 41} — found sensitive in ablations
Mamba 1D conv	BF16	All layers
Router gates	FP32	Routing precision must not degrade
Embeddings & lm_head	BF16	Not quantized
KV cache	FP8	All 6 attention layers

Calibration

• Dataset: 4,000 samples from nvidia/Nemotron-Cascade-2-SFT-Data
• Domain mix: math (1000), swe (900), terminal_agent (500), science (500), chat (400), conversational_agent (300), instruction_following (300), safety (100)
• Sequence length: Up to 12,288 tokens (no padding, natural length per sample)

Usage

SGLang

python -m sglang.launch_server \
    --model chankhavu/Nemotron-Cascade-2-30B-A3B-FP8 \
    --trust-remote-code \
    --tool-call-parser qwen3_coder \
    --reasoning-parser nano_v3

vLLM

vllm serve chankhavu/Nemotron-Cascade-2-30B-A3B-FP8 \
    --mamba_ssm_cache_dtype float32 \
    --max-model-len 262144 \
    --trust-remote-code \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --reasoning-parser nemotron_v3

Acknowledgments

• Quantization recipe based on the Nemotron 3 Nano Technical Report
• Quantized with NVIDIA ModelOpt