PreDoctor/DeepSeek-R1-0528-Qwen3-8B-NVFP4-W4A16-PreTensor

R1 8B NVFP4 W4A16 PreTensor

This repository contains the reasoning-tuned R1 8B checkpoint packaged for PreTensor's local NVIDIA CUDA runtime.

• Language weights: packed FP4 E2M1, group size 16
• Language weight scales: FP8 E4M3 plus FP32 tensor scales
• Language activations and KV cache: BF16 (W4A16)
• Input and output: text only; vision and audio encoders are not included
• Runtime path: native NVIDIA CUDA host; WebGPU payloads are not included

The files are valid sharded safetensors, but their packed tensor contract requires the PreTensor loader. See config.json for the storage contract and pretensor-nvfp4-weights.json for the tensor inventory and file SHA-256 values.

Model identity and size

• Architecture: Qwen3ForCausalLM, 36 transformer layers
• Quantized tensor payload: 6,396,931,056 bytes (5.958 GiB)
• Configured context limit: 131,072 tokens

The payload size is compressed on-disk storage, not a parameter count or a VRAM requirement. Runtime buffers and the resident KV cache require additional memory.

Long-context prefill validation

Warm native measurements on NVIDIA RTX PRO 6000 Blackwell Max-Q (CUDA device 0, driver 620.12). Each result is the median of 3 complete session runs using the released PreTensor 0.1.23 native host.

Input tokens	Prefill	Native wall time
4,096	6,743 tok/s	607.4 ms
16,384	3,347 tok/s	4,894.9 ms

These figures validate this PreTensor implementation on the stated machine; they are not general hardware-independent model benchmarks.

Package contents

• Stored tensors: 903
• Quantized projection matrices: 252
• Tensor payload: 6,396,931,056 bytes (5.958 GiB)
• Safetensors shards: 4

All four weight shards and model.safetensors.index.json are required. The tokenizer, chat template, configuration and MIT license are distributed with the checkpoint. The source model is DeepSeek-R1-0528-Qwen3-8B.