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.
