StargazerLabs/Qwen3.8-23B-Mini-Me-bf16

Qwen3.8-23B-Mini-Me (bf16)

This model has been through a battery of personal probes rather than standard benchmarks; there are no standard benchmark metrics to report yet.

During internal use, it has held for coding and agentic work — long multi-turn conversations, tool calling, instruction retention past its nominal context — while landing slightly short of the parent across most things.

Expect it to be smaller, faster, and just a little bit less smart than Qwen3.8-27B, not a different model.

Architecture: Same as the original Qwen3.8-27B minus 12 layers: 52 layers, vision tower intact and untouched

Layers 12–15, 24–27, and 36–39 were all removed based on lesion probing across different combinations of depth prunes; these were the most favorable combination of layers to remove based on internal tests