An MLX affine 4-bit (group size 64) build of harness-hub/Qwen3.8-27B-pi-adapted. That LoRA tunes Qwen3.8-27B for the pi coding agent.
This repository is a drop-in replacement for TensorFold/Qwen3.8-27B-MLX-4bit. The shard layout, tensor names, config, tokenizer and chat template are the same. Only the 256 projections that the LoRA targets have different bytes.
Qwen/Qwen3.8-27B at revision 37561816.harness-hub/Qwen3.8-27B-pi-adapted at revision e786f849 (r 32, alpha 64, scaling 2.0): W' = W + 2.0 * B @ A.gate_proj, up_proj and down_proj in all 64 layers, and q_proj, k_proj, v_proj and o_proj in the 16 full-attention layers (3, 7, ..., 63).TensorFold/Qwen3.8-27B-MLX-4bit at revision 22d8d538. This includes the vision tower, embeddings and linear-attention layers.The quantizer reproduces the rounding of the MLX 0.32.4 Metal kernel exactly. The MLX CPU quantizer rounds ties to even and gives slightly different bytes. As a gate, each of the 256 target projections was first re-quantized from the unmerged BF16 weights. The result had to match the TensorFold bytes exactly, and it did for all 256 (weights, scales and biases).
This checkpoint was checked only under the TensorFold engine, in agent workloads. The format is the same as TensorFold/Qwen3.8-27B-MLX-4bit, so it should also load in mlx-vlm. That path was not tested:
pip install mlx-vlm
python -m mlx_vlm.generate --model tozuo/Qwen3.8-27B-pi-adapted-MLX-4bit \
--prompt "Write a bash one-liner that counts lines in all .py files." \
--max-tokens 512
The adapter card recommends thinking off. The chat template enables thinking by default. To turn it off, send chat_template_kwargs: {"enable_thinking": false} from an OpenAI-compatible server.
The checkpoint needs about 16 GB of unified memory for the weights, plus the KV cache.
As required by the Apache License 2.0, section 4(b):
TensorFold/Qwen3.8-27B-MLX-4bit. They are the result of merging the harness-hub LoRA and re-quantizing.TensorFold/Qwen3.8-27B-MLX-4bit, except this README.Apache License 2.0, the same license as Qwen/Qwen3.8-27B, harness-hub/Qwen3.8-27B-pi-adapted and TensorFold/Qwen3.8-27B-MLX-4bit. See LICENSE.
model-00003-of-00003.safetensors 5.4 GBmodel-00002-of-00003.safetensors 5.4 GBmodel-00001-of-00003.safetensors 5.3 GBQwen/Qwen3.8-27B