Qwen3.5-9B INT4 OpenVINO (text generation)
Maintained by AtomGradient / 质子梯度（北京）科技有限公司.
The text-generation files of the OpenVINO™ IR export of Qwen3.5-9B with INT4 weight compression published by the OpenVINO organisation, pinned at one revision, with one chat-template change. AtomGradient uses these files in its local speech delivery; the runtime is supplied separately and is not part of this repository.
Model specifications
| Item | Specification |
|---|---|
| Base model | Qwen/Qwen3.5-9B |
| Model size class | 9B |
| Weights | INT4 asymmetric, group size 128, ratio 1.0 (NNCF compress_weights, as published upstream) |
| Format | OpenVINO IR: language model, text embeddings, tokenizer and detokenizer; no vision files |
| Compatibility | OpenVINO 2026.2.0 and later, as stated by the upstream export |
Provenance and changes
| Item | Value |
|---|---|
| Upstream repository | OpenVINO/Qwen3.5-9B-int4-ov |
| Upstream revision | 3f31fc1609edb0cbc8b131bc4c291f1b15cb0676 |
| Model weights and configuration | Byte-identical to the upstream revision for every file listed in SOURCE.json except chat_template.jinja |
chat_template.jinja |
Sets enable_thinking = false at the top of the template so the model answers without a thinking block; the rest of the template is unchanged (PROVENANCE.json records the original template digest) |
| Files not included | The upstream vision encoder, merger and position files |
SOURCE.json lists every file with its size and SHA-256; ARTIFACT.sha256 can be checked with sha256sum -c.
Scope
- Text-only generation; thinking is disabled by the template. Vision inputs are not supported by this file set.
- The model can state facts incorrectly or choose a tool wrongly; verify important information.
License
Apache-2.0, inherited from the base model and the upstream export; see LICENSE.txt.
- Downloads last month
- -