SemIf Qwen3.5-4B — NovelAide MLX mirror
This repository mirrors the text-only affine 4-bit/group-64 export from
vinci00/semif-qwen3.5-4b-mlx-4bit,
fixed at revision 4ef6ec3e6e41f95ce6a6cf6deb0e6184753dddde.
The installation files are unchanged copies under assets/2026.10.05-mlx-4bit/.
The source checkpoint is Qwen/Qwen3.5-4B,
fixed at 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. No fine-tuning or adapter is included.
The export's decision_manifest.json records its source and conversion hashes.
SemIf by TheoLeeCJ supplies the criterion/evidence prompt and the softmax over the supplied answer letters' final-position logits. That decision readout does not generate answer text. NovelAide executes it in its native Swift MLX runtime on Apple Silicon; these assets contain weights and metadata, with no Python service or runtime executable. Generic text generation does not reproduce the decision readout.
The package is for local criterion-based selection and classification. Its probabilities are conditional on the supplied options and are not guarantees of truth. This mirror does not claim new training, broader language quality, long-context quality, or a GGUF implementation.
Attribution and license
Qwen3.5-4B is provided by the Qwen team under Apache-2.0; see LICENSE-QWEN.txt. SemIf's prompt and readout implementation are MIT licensed; see LICENSE-SEMIF.txt. The quantized export and its verification package are by vinci00/VinciGit00, using Apple's MLX-LM. Both upstream license notices are also included in the installation file set. The original benchmark used JevBench; no benchmark dataset is redistributed in this mirror.
4-bit