Qwen3.8-23B-Mini-Me: Depth-pruned 27B model runs faster, no fine-tuning
Layer removal cuts 4.3B params, boosting speed for coding and agent tasks without retraining.
Peplo1214, a solo developer, shared Qwen3.8-23B-Mini-Me on Reddit, a depth-pruned variant of Alibaba's Qwen3.8-27B. Depth pruning works by removing entire transformer layers rather than individual weights, shrinking the model from 27B to roughly 22.7B parameters. Crucially, the creator did not fine-tune the model after pruning—just strategically removed layers, which preserves most reasoning capabilities while cutting the footprint by ~16%. The model is available in bf16, q8, and q4 quantizations, though only MLX (Apple Silicon) versions are currently published.
Early reports suggest it performs well on standard coding problems, agentic workflows, and multi-turn conversations. However, the creator notes it struggles with edge cases and underspecified prompts where the original 27B model's extra capacity helps infer intent. They also recommend using the exact same chat settings as Qwen3.8-27B to avoid looping. With no official benchmarks, peplo1214 is transparent that this is a smaller version that's "slightly worse at some things but has a smaller footprint and runs faster," inviting the community to stress-test it for production use.
- Depth-pruned from 27B to ~22.7B parameters by removing layers, with zero fine-tuning.
- Available in bf16, q8, and q4 formats, but currently MLX-only for Apple Silicon.
- Strong on coding and agentic tasks, but weaker on underspecified prompts and edge cases.
Why It Matters
Demonstrates that layer pruning can shrink dense LLMs cheaply, enabling faster, lower-cost deployment on consumer hardware.