Tiny3.5
An attempt to compress Qwen3.5 into 500M and 1.5B parameters.
What is this?
Tiny3.5 is my community effort to create tiny and more efficient versions of Qwen3.5. The strengths of Tiny3.5 include very low inference latency, minimal overthinking, and being able to run on much weaker hardware. However, it's important to realize that Tiny3.5 is sub-2B parameters. Don't expect a 99% score on every single benchmark.
How is this better than Qwen3.5?
Tiny3.5 uses many techniques to produce better efficiency than Qwen3.5 in many scenarios. We use multi-shot distillation to filter out pointless reasoning loops and improve the overall quality of responses.
Can I create my own model using the Tiny3.5 dataset?
Absolutely! Our distillation dataset is open-source, and the code used to create it alongside a copy of the dataset is available on our GitHub: https://github.com/reecdev/tiny3.5
- Downloads last month
- 67
16-bit