QSIV
Copyright 2026 the QSIV authors.

This repository contains model weights derived from Qwen3.5-0.8B-Base (https://huggingface.co/Qwen/Qwen3.5-0.8B-Base),
Copyright the Qwen team (Alibaba Cloud), licensed under the Apache License, Version 2.0. Modifications: execution
surgery (split12 evidence-memory layout, windowed state attention, per-question isolation), typed-decision fine-tuning,
FP8 quantisation-aware training and post-hoc temperature calibration.

The included tokenizer files are those of Qwen3.5-0.8B-Base (Apache-2.0).

The inference kernels implement Gated DeltaNet (Yang et al.) in Triton, following the flash-linear-attention algorithms
(fla-org, MIT). The optional bf16 reference path calls flash-linear-attention.

The model was fine-tuned on third-party datasets that keep their own licences. No dataset content is distributed in this
repository.