| Wren |
| Copyright 2026 ohmychemo (modifications) |
|
|
| Wren is a derivative work of Swift-Qwen3.8-Flash-Next (Swift 1.5) by UkisAI, |
| https://huggingface.co/ukisai/Swift1.5-Qwen3.8-Flash-Next |
| Copyright 2026 UkisAI, Swift Open License v1.0 (see LICENSE), |
| which is itself a derivative work of Qwen3.8-Flash-Next, |
| https://huggingface.co/Qwen/Qwen3.8-Flash-Next |
| Copyright (c) 2026 Qwen, Qwen Community License 1.0 (see LICENSE-QWEN). |
|
|
| "Swift" and "UkisAI" are trademarks of UkisAI. They are used here only to |
| describe the origin of this work. Wren is not affiliated with or endorsed by |
| UkisAI or Qwen. |
|
|
| Changes made by ohmychemo: |
| - Routed experts: 256 of the 512 experts per layer were removed (staged |
| pruning 100 -> 75 -> 60 -> 50%, maximin over REAP saliency on 13 |
| calibration areas); the router rows were reduced accordingly. The MTP head |
| was removed. |
| - Router and normalisation weights were retrained by logit distillation from |
| Swift's own top-20 log-probabilities ("healing"). |
| - config.json updated for the reduced expert count; kept_experts.json and |
| training/ (logs, masks, evaluations) added; README.md replaced. |
| - GGUF files (ohmychemo/Wren-GGUF): quantised with llama.cpp using imatrix |
| data from Swift's own outputs. Wren_T3 and Wren_T4-* store each layer's |
| experts in two groups at different quantisation types (tiered-experts |
| llama.cpp fork); router rows are permuted to match. Evaluation logs, |
| allocation files and the research paper (paper/) added. |
| - All other files are unmodified from the Swift release. |
|
|
| |
| Original NOTICE of Swift-Qwen3.8-Flash-Next, reproduced as required: |
|
|
| Swift-Qwen3.8-Flash-Next |
| Copyright 2026 UkisAI |
|
|
| UkisAI's contribution (the "Swift Contribution") is licensed under the |
| Swift Open License v1.0. See LICENSE. |
|
|
| This model is a derivative work of Qwen3.8-Flash-Next |
| https://huggingface.co/Qwen/Qwen3.8-Flash-Next |
| Copyright (c) 2026 Qwen |
| Licensed under the Qwen Community License 1.0. See LICENSE-QWEN. |
|
|
| Changes made by UkisAI: |
| - model-*.safetensors: model weights were modified by UkisAI through fine-tuning. |
| - tokenizer.json, tokenizer_config.json: re-serialized from UkisAI's training |
| base. Vocabulary and merges are unchanged; the file format, some metadata |
| fields and the pre-tokenizer pattern differ from the Base Model files. |
| - README.md: replaced. LICENSE (Swift Open License v1.0), NOTICE, |
| benchmarks/ and ukisai-banner.png added. |
| - The Base Model's license text was moved from LICENSE to LICENSE-QWEN. |
| - All other files (config.json, generation_config.json, chat_template.jinja, |
| model.safetensors.index.json, vocab.json, merges.txt, |
| preprocessor_config.json, video_preprocessor_config.json) are unmodified |
| from Qwen3.8-Flash-Next and remain under the Qwen Community License 1.0. |
|
|
| GSQ-RCO GGUF release changes by UkisAI: |
| - Swift Flash Next weights converted to mixed-precision GGUF, with |
| reused ISTA-DASLab allocation profiles and Swift-specific refinement. |
| - Weights split into two GGUF shards per tier; BF16 vision projector included. |
| - Quantization metadata, evaluation records and an adapted README added. |
|
|
