harness-hub/Qwen3.8-27B-pi-adapted (a LoRA trained with RL for the pi coding agent) merged into Qwen/Qwen3.8-27B and quantized to INT8 in the format of bpx03/Qwen3.8-27B-INT8-QEngine, so it runs directly on QEngine (1.3.6) on 2-3x RTX 3090.
All credit for the adapter and its results goes to harness-hub; for the base model to the Qwen team; for the engine, the INT8 format and the conversion tool to bpx03. This repo only does the merge and the conversion.
Qwen/Qwen3.8-27B @ 37561816e66b5b9153b1b64c10968c18c1e6f4a0 (BF16).harness-hub/Qwen3.8-27B-pi-adapted @ e786f8491742268f9b2bc729697b13bfac26d4ad
(LoRA r=32, alpha=64; q/k/v/o of the 16 full-attention layers and gate/up/down of all 64 layers; 256 A/B pairs).W' = W + (alpha/r) · B·A. Every LoRA pair mapped to exactly one base weight.444b2e49) tools/quant_int8.py,
FP8 KV scales taken from QEngine's 27B INT4 checkpoint (--kv-from), as the QEngine README describes.token_rank.bin is copied unchanged from bpx03/Qwen3.8-27B-INT8-QEngine (the tool does not produce it).Merging before quantizing matters: adding a LoRA delta to already-INT8 weights would round most of it away.
Sanity check against bpx03/Qwen3.8-27B-INT8-QEngine (same pipeline, no adapter): embeddings, lm_head, norms, all 48
DeltaNet layers, KV scales, the vision tower and model_mtp.safetensors are byte-identical; only the adapted
projections differ, each changed element by exactly one INT8 step (gate 5.6%, up 5.5%, q/k 3.4%, o 2.8%, v 2.4%,
down 2.1% of elements). The MTP head is the stock one (the adapter does not touch it).
The tokenizer files are the adapter's (newer pre-tokenizer rule for combining marks than the bpx03 INT8 repo; same vocabulary). The chat template is byte-identical to the base model's.
hf download <this repo> --local-dir models/Qwen3.8-27B-pi-INT8
# Docker: -e QE_MODEL=/app/models/Qwen3.8-27B-pi-INT8
Serve it like the bpx03 INT8 checkpoint. Match how the adapter was trained and evaluated:
ENABLE_THINKING=falseQE_PRESENCE_PENALTY=0 (QEngine 1.3.6 defaults to 1.5 for non-thinking requests)--loop-guard-think 0 (otherwise QEngine can force a reasoning turn when an agent repeats itself)"api": "openai-completions" (not the Responses API)| this model | base INT8 | |
|---|---|---|
| decode, prose / code, tok/s (single stream, 500 tokens) | 117.9 / 240.4 | 122.5 / 243.9 |
| DFlash2 mean accepted length, prose / code | 2.43 / 5.04 | – |
Tool calls (tool_choice auto / required / none / named), a 32K needle and basic arithmetic passed.
**The agent benchmark gains reported on the adapter card (Terminal-Bench 2.1 39.5% → 55.8%, SWE-bench Pro hard-51 29.4% → 62.7%, under pi) were measured on the BF16 adapter; see the small INT8 check below. **
Terminal-Bench 2.1, 20 tasks (seed 20261117, sampled from all 89 tasks; the card's 18 checkpoint-selection
tasks are not published, so they could not be excluded), Harbor 0.26.0 with its pi agent, pi 1.5.2,
openai-completions, thinking off, temperature 0.8, presence penalty 0, 100 turns, 1 h per task,
1 attempt. Both arms are INT8 on QEngine 1.3.6; the only difference is the merged adapter.
| this model | base INT8 (bpx03) | |
|---|---|---|
| tasks solved | 10 / 20 (50%) | 8 / 20 (40%) |
| tool calls per task, mean | 24.1 | 50.1 |
| runs hitting the 100-turn limit | 2 | 8 |
Paired: +3 newly solved, −1 lost (exact sign test p = 0.63). The direction and most of the tool-call reduction (−52% vs the card's −62%) survive the INT8 conversion, but n = 20 is too small to confirm the card's +16 points. No malformed tool calls. About 1% of replies start with a short reasoning-style preamble ending in ``; with vLLM's reasoning parser this is stripped, QEngine with thinking off passes it through.
Apache-2.0, as the base model, the adapter and the bpx03 INT8 checkpoint.
model.safetensors 31 GBmodel_mtp.safetensors 849 MBtoken_rank.bin 1 MBharness-hub/Qwen3.8-27B-pi-adapted