Qwen3.8-27B-pi-adapted-bf16 is a standalone BF16 checkpoint made by merging the harness-hub/Qwen3.8-27B-pi-adapted LoRA adapter into Qwen/Qwen3.8-27B, so it serves on any engine without LoRA support. The adapter (r = 32, alpha = 64, on the q, k, v, o, gate, up and down projections) was trained with AsyncGRPO reinforcement learning in Harbor/OpenEnv environments to make the model work as the backbone of the pi coding agent on software-engineering and terminal tasks, with thinking off as in training. The merge is the standard
W + (alpha/r)·B@Aupdate, applied per tensor in fp32 and cast back to BF16, with the base shard layout, tensor names and all MTP tensors preserved. According to the adapter's card, under pi it raises Terminal-Bench 2.1 from 39.5% to 55.8% and SWE-bench Pro (hard 51) from 29.4% to 62.7%, while using 62% and 55% fewer tool calls per task. These figures were measured for the adapter on the BF16 base, not re-evaluated on this merged checkpoint.
[!NOTE] MTP — Speculative Decoding: Yes
[!NOTE] Multi-Token Prediction (MTP) GGUF is a specialized GGUF model file format extension that integrates speculative decoding directly into the model weights to significantly accelerate local inference. Unlike traditional speculative decoding which requires a separate, smaller "draft" model, MTP GGUF files include additional output heads within the main model architecture that predict multiple future tokens in a single forward pass.
| File Name | Quant Type | File Size | File Link | Description |
|---|---|---|---|---|
| Qwen3.8-27B-pi-adapted.Q4_K_M.gguf | Q4_K_M | 17 GB | Link | Good quality, default size for most use cases, recommended. |
| Qwen3.8-27B-pi-adapted.Q4_K_S.gguf | Q4_K_S | 15.8 GB | Link | Slightly lower quality with more space savings, recommended. |
| Qwen3.8-27B-pi-adapted.Q5_K_M.gguf | Q5_K_M | 19.6 GB | Link | High quality, recommended. |
| Qwen3.8-27B-pi-adapted.Q5_K_S.gguf | Q5_K_S | 19 GB | Link | High quality with slight space savings, recommended. |
| Qwen3.8-27B-pi-adapted.mmproj-bf16.gguf | mmproj-bf16 | 931 MB | Link | Multimodal projection file in BF16 format. Used for vision/language capabilities. |
LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp
Quantized using llama.cpp by prithivMLmods.
harness-hub/Qwen3.8-27B-pi-adapted