Gemma 4 12B W4A16 PreTensor

This repository contains the instruction-tuned Gemma 4 12B checkpoint packaged for PreTensor's local NVIDIA CUDA runtime.

Language weights: compressed-tensors packed INT4, symmetric group size 32
Language activations and KV cache: BF16
Vision input: source BF16 tensors, copied without quantization
Audio input: native 16 kHz waveform projection in BF16 (12B only)
Tensor payloads: copied from the source checkpoint without requantization
Runtime path: native NVIDIA CUDA host; WebGPU payloads are not included

The files are valid sharded safetensors. See pretensor_config.json for the storage contract, source revision, component inventory, and per-shard SHA-256.

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	8,618 tok/s	475.3 ms
16,384	6,335 tok/s	2,586.3 ms

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

Package contents

Language tensors: 1,323
Vision tensors: 10
Audio tensors: 1
Tensor payload: 10,264,057,056 bytes
Safetensors shards: 3

The tokenizer, processor configuration, chat template, license, and notice are distributed with the checkpoint. Source revision: 1d2c2d7f2466070e69d6fb3fd5ce9a7d75f2f6ee.