Qwen3.8-27B NVFP4 W4A16 Multimodal for PreTensor
This is the complete local PreTensor distribution for Qwen3.8-27B. It contains
the language model in PreTensor's pre-quantized NVFP4 W4A16 format and the
original vision encoder and multimodal merger in BF16.
The model runs through the PreTensor Chrome extension and its local NVIDIA CUDA
host. It is not a WebGPU model and does not use a browser WebGPU inference
runtime.
- Language weights: packed FP4 E2M1, group size 16
- Language weight scales: FP8 E4M3FN plus FP32 global scales
- Language activations: BF16 (W4A16)
- Vision encoder and multimodal merger: original BF16, not quantized
- Text tensors copied byte-for-byte from the verified PreTensor language checkpoint
- Vision tensors copied byte-for-byte from the source model
The files are valid sharded safetensors, but the packed tensor names and fused
Gate/Up layout require the PreTensor loader. Stock Transformers does not load
this pre-quantized checkpoint directly. This limitation is explicit so that
Transformers does not silently interpret packed FP4 bytes as ordinary weights.
See pretensor_config.json for the exact storage contract and shard hashes.
Model identity and size
- Model architecture: Qwen3.8-27B
- Source parameter count: 27,781,427,952 parameters (27.8B)
- Quantized tensor payload: 17,881,899,428 bytes (16.654 GiB)
The 16.654 GiB figure is compressed on-disk storage, not a 16B parameter
model. NVFP4 weights are packed into byte arrays, so counting stored array
elements does not recover the source model's parameter count. The exact source
parameter count is recorded as metadata.total_parameters in
model.safetensors.index.json.
Long-context prefill validation
Warm native measurements on NVIDIA RTX PRO 6000 Blackwell Max-Q (CUDA device 0,
driver 620.12), after model/graph and exact KV-capacity warm-up. Each result is
the median of 3 runs.
| Input tokens | Prefill | Native wall time |
|---|
| 4,096 | 4,155 tok/s | 1,020.4 ms |
| 16,384 | 3,707 tok/s | 4,460.9 ms |
These figures validate this PreTensor implementation on the stated machine;
they are not general hardware-independent model benchmarks.
Package contents
- Safetensors shards: 5
- Total tensors: 1,986
- Text tensors: 1,653
- Vision tensors: 333
- Quantized tensor payload: 17,881,899,428 bytes (16.654 GiB; not parameter count)
All five model-0000x-of-00005.safetensors files and
model.safetensors.index.json are required. The language tensors span shards
1-4; the vision tensors span shards 4-5.
The tokenizer, processor configuration, chat template, and license are copied
from the source Qwen/Qwen3.8-27B model revision recorded in
pretensor_config.json.
