phaedawg's picture
Upload NVFP4 quantized Behemoth-R1-123B-v2
6ad90cc verified
metadata
license: other
base_model: TheDrummer/Behemoth-R1-123B-v2
tags:
  - nvfp4
  - modelopt
  - quantized
  - blackwell
  - b200
library_name: transformers

Behemoth-R1-V2 ModelOpt NVFP4

NVFP4 quantized version of TheDrummer/Behemoth-R1-123B-v2 using NVIDIA Model Optimizer.

Quantization Details

Property Value
Original Model TheDrummer/Behemoth-R1-123B-v2
Quantization NVFP4 (FP4 weights, FP16 activations)
Method NVIDIA ModelOpt PTQ
Calibration Samples 512
Max Sequence Length 4096

Hardware Requirements

  • Optimal: NVIDIA Blackwell GPUs (B100, B200, RTX PRO 6000 Blackwell)
  • Compatible: Hopper/Ampere (will use weight-only mode)

Usage with vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="TheHouseOfTheDude/Behemoth-R1-V2_ModelOpt-NVFP4",
    quantization="modelopt",
    trust_remote_code=True,
)

sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=512)
outputs = llm.generate(["Write a story about..."], sampling_params)
print(outputs[0].outputs[0].text)

Chat Template

Uses Mistral v7 (Non-Tekken) format. See the original model card for usage details.

Credits

  • Original Model: TheDrummer
  • Quantization: TheHouseOfTheDude
  • Quantization Framework: NVIDIA ModelOpt