Qwen 2.8 (1.5B)
a compact reasoning model with mode switching
2 min readupdated
VoidOSS/Qwen-2.8-1.5B is a lightweight language model based on Qwen/Qwen2.5-1.5B (a 1.5-billion-parameter base model from Alibaba Cloud), fine-tuned on DeepSeek-R1 distillation data. Its standout feature is a built-in step-by-step reasoning mechanism ... and support for switching between Thinking mode (deep analysis) and Fast mode (quick replies without reasoning). It is optimized to run on modest hardware like GPUs with 8 GB of VRAM and significantly outperforms the base version on math and logic tasks. This release is useful for developers and hobbyists building smart local assistants.
Key features
- Compact size: 1.5 billion parameters for efficient inference on home PCs and entry-level GPUs.
- Dual-mode operation: supports weight-level switching between reasoning mode and fast responses with 100% bypass reliability.
- Improved reasoning: fine-tuned on DeepSeek-R1 teacher data to tackle complex multi-step tasks.
What you can do
- Run the model locally on an 8 GB VRAM GPU to solve math and coding problems offline.
- Integrate the model into a personal assistant bot, toggling the reasoning block via prompt prefixes.
Example
To enable Fast Mode without generating a chain of thought, pass the prefix <|im_start|>assistant
at the start of the response; for reasoning mode, use the standard system prompt.
Read more
DetailsGSM8K (math): score jumped from 12.0% to 40.0% (a 3.33x increase).
Logic and coding: Logic & Cognitive Traps benchmark improved from 30.0% to 70.0%.
Testing: benchmarks were measured on an RTX 4060 Ti (8 GB VRAM).
Training: replay data mixing combined with distillation from DeepSeek-R1.










Comments