Qwen3.8-27B-pi-adapted Code and Terminal Task LoRA Adapter: 4GB+ VRAM

November 20, 2026

Qwen3.8-27B-pi-adapted Code and Terminal Task LoRA Adapter: 4GB+ VRAM

At a Glance

Item Value
Repository harness-hub/Qwen3.8-27B-pi-adapted
Family guide Qwen3.8 guide (11 articles)
Publisher guide Alibaba (Qwen): models and licenses
Published 2026-11-16
License apache-2.0
Formats safetensors
Source type Unverified (not confirmed by a primary source)

Values determined by this site’s code when the information was collected. Dates are JST.

What we checked ourselves

  • It needs about 21% fewer tokens than the Qwen3 tokenizer for the same Japanese text, so about 1.26× as much Japanese fits in the same context length, and generation is faster per character.

Details and conditions are in “Our Own Measurements” below.

Overview

[Unverified] This information is based on the model page on Hugging Face and has not been confirmed as an official announcement.

It is reported that harness-hub has released the LoRA adapter “harness-hub/Qwen3.8-27B-pi-adapted" for Qwen/Qwen3.8-27B. According to the model card, it is a model trained via reinforcement learning for software engineering and terminal operation tasks, designed to run behind the coding agent “pi". It is not a standalone model, but an adapter used on top of Qwen/Qwen3.8-27B.

Specifications

  • Type: LoRA adapter (PEFT). Base model is Qwen/Qwen3.8-27B
  • Base model parameters: 27B (Dense model, as stated in the original model card)
  • LoRA configuration: r = 32, alpha = 64, targets q, k, v, o, gate, up, down projections
  • Training method: AsyncGRPO (TRL, experimental). Reward is task verifier correctness plus a tool-efficiency bonus (weight 0.1)
  • Training scale: Checkpoint at step 160 (stopped at step 191 out of planned 1000 steps)
  • Base model context length: Native 262,144 tokens, expandable up to 1,000,000 tokens. Evaluation and serving settings in the card specify --max-model-len 131072

Performance

Below are figures published by the publisher itself in the model card and have not been independently verified. Evaluations were conducted under the pi harness, 1 trial per task, 100-step limit, text only, temperature 0.8. It is explicitly noted that these cannot be compared with the base model card’s figures (using harnesses like Terminus, Claude Code, etc.).

1st table compares the base model with thinking (xhigh). The original table lists the column name as “Qwen3.8-27B-opencode-adapted", but the numbers match those in the text (64.7%, 56.6%) and are considered to be values for this adapter. Column names are kept as in the original.

→ Scroll horizontally to see all columns

Benchmark Qwen3.8-27B Qwen3.8-27B, xhigh thinking Qwen3.8-27B-opencode-adapted Qwen3.8-27B-opencode-adapted, xhigh thinking
SWE-bench Pro (hard 51) 39.2% 64.7% 64.7% 66.7%
Terminal-Bench 2.1 41.0% 54.2% 56.6% 56.6%

A paired comparison with the base model (thinking off) is as follows:

→ Scroll horizontally to see all columns

Benchmark Tasks Qwen3.8-27B Qwen3.8-27B-pi-adapted (step 160) Difference (points) Newly solved / lost
SWE-bench Pro (hard 51) 51 39.2% 64.7% +25.5 ± 14.2 15 / 2 (p = 0.0023)
Terminal-Bench 2.1 86 39.5% 55.8% +16.3 ± 9.1 16 / 2 (p = 0.0013)
Terminal-Bench 2.1, the 68 tasks not used to pick the checkpoint 68 39.7% 55.9% +16.2 ± 10.6 13 / 2 (p = 0.0074)

From these tables, the adapter without thinking significantly outperforms the base model by +25.5 points on SWE-bench Pro (whether real repository issues can be fixed; higher is better) and +16.3 points on Terminal-Bench 2.1 (whether terminal operations can be completed; higher is better). It ties with the base model with xhigh thinking at 64.7% on SWE-bench Pro, and is closely matched at 56.6% vs 54.2% on Terminal-Bench 2.1. On the other hand, enabling thinking on the adapter showed no improvement (66.7% on SWE-bench Pro, 56.6% on Terminal-Bench 2.1), and the card recommends “using it with thinking off". Median generated tokens per task are 7.9k for the adapter versus 34.4k for the thinking-enabled base model on SWE-bench Pro, and 3.7k versus 16.5k on Terminal-Bench 2.1. Note that evaluation is limited to the 51 hard tasks of SWE-bench Pro, and Terminal-Bench 2.1 includes tasks used for checkpoint selection. Additionally, it is stated that the GGUF version has not been re-evaluated.

Strengths and Use Cases

Note that this has not been officially confirmed, and the following is based on the model card descriptions.

This adapter is specialized for software engineering and terminal operation tasks, designed to run behind the “pi" coding agent. According to the card, on actual SWE-bench Pro tasks, while the base model repeatedly searched and re-ran snippets until hitting the 100-step limit, the adapter identified the function, edited it, ran tests, and fixed the final case in just 8 tool calls. For tasks solved by both, it reportedly requires less than half the turns of the base model.

Suitable use cases include:

  • Autonomous tasks involving bug investigation and fixing within repositories, and running tests
  • Tasks to complete operations by executing commands in a terminal
  • Agent operations keeping token counts low without thinking (the card recommends keeping thinking disabled)

There are also points to note:

  • Evaluations are text-only, and no vision tower is used. Although the base model can handle images and video, image input with this adapter has not been verified
  • Evaluations assume the pi harness. According to the card, the API must specify "api": "openai-completions". It has neither been trained nor evaluated with the Responses API
  • The adapter only works on top of Qwen/Qwen3.8-27B and does not operate with other base models
  • Improvements in general conversation, mathematics, or scientific reasoning cannot be gleaned from the documentation

How It Differs from Similar Models

In our site’s previous coverage of the Qwen3.8 family, there is “Qwen3.8-27B-DSpine-b16" Speculative Decoding Draft Model: 8GB+ VRAM Required. This is a draft model for speculative decoding, used to speed up generation in Qwen3.8-27B. This adapter differs in its objective, using reinforcement learning to alter the base model’s behavior specifically for pi-related tasks.

Additionally, the Decision-Making Model “vllm-d3-GGUF" Series Released is a GGUF converted version for decision-making tasks returning judgment results without text generation, targeting a different use case than this coding agent adapter.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 27.8B parameters (taken from the base model Qwen/Qwen3.8-27B)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) F16 0.3GB 0.4GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build harness-hub/Qwen3.8-27B-pi-adapted-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but harness-hub/Qwen3.8-27B-pi-adapted-GGUF provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Values we measured ourselves on our server (no GPU) by actually reading and running this model’s files — not figures copied from the model card. How we measure, and the results for every model: Our Measurements.

Japanese Token Efficiency

Tokenizer Tokens per 1,000 Japanese characters Ratio to the same text in English
This model 546 0.99×
Qwen3 688 1.26×
Llama 3.2 744 1.36×
Gemma 3 564 1.03×
gpt-oss 795 1.45×
LLM-jp-3 497 0.85×

It needs about 21% fewer tokens than the Qwen3 tokenizer for the same Japanese text, so about 1.26× as much Japanese fits in the same context length, and generation is faster per character.

Counted with the tokenizer.json of harness-hub/Qwen3.8-27B-pi-adapted on a fixed text we wrote ourselves (876 Japanese characters across news, conversation, technical docs, a formal email, travel writing and a recipe) and its English translation. Fewer tokens mean more Japanese fits in the context window.

Inside the GGUF File

File: Qwen3.8-27B-pi-lora-F16.gguf (0.30GB, F16). We read only the header (metadata) of the file, not the weights.

Item Value
Architecture (as named in the GGUF) qwen35
Chat template Not included
imatrix Not recorded in the file
Weight types (share of parameters) F16 100.0%
Average bits per weight 16.00 bits

The quant name in the file name describes the file as a whole; in practice layers mix several types. The average bits per weight is the measured file data divided by the number of weights.

Recent Models in the Same Size Class

Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
ggml-org/Muse-Glimmer-30B-GGUF 29.8B 24GB apache-2.0 Muse-Glimmer-30B-GGUF 30B Multimodal Local Agent Model: 24GB+ VRAM (2026-11-17)
Cloudflare/clef-omni 35.3B 80GB apache-2.0 clef-omni 30B Multimodal Decision Model: 80GB+ VRAM (2026-11-17)
inclusionAI/ZwZ2-27B-GGUF 27.4B 24GB — ZwZ2-27B-GGUF Vision-Language Model: 24GB+ VRAM (2026-11-17)
ggml-org/OpenJev-GGUF 27.4B 24GB cc-by-nc-4.0 OpenJev-GGUF 27.4B Decision-Making Specialized Model: 24GB+ VRAM (2026-10-01)

How to Get It

Distributed in safetensors format (PEFT LoRA adapter). The license is apache-2.0, with no gated agreement required on Hugging Face. The base model Qwen/Qwen3.8-27B is required separately.

For vLLM, the card suggests the following command for evaluation configuration:

vllm serve Qwen/Qwen3.8-27B \
    --enable-lora --max-lora-rank 32 \
    --lora-modules qwen3.8-pi=harness-hub/Qwen3.8-27B-pi-adapted \
    --language-model-only --enable-prefix-caching --mamba-cache-mode align \
    --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 \
    --default-chat-template-kwargs '{"enable_thinking": false}' \
    --max-model-len 131072

Installation of pi is as follows (Node.js 22.19 or higher):

npm install -g --ignore-scripts @earendil-works/pi-coding-agent

When using with llama.cpp, layer the GGUF LoRA adapter (harness-hub/Qwen3.8-27B-pi-adapted-GGUF) onto the base model’s GGUF.

hf download ggml-org/Qwen3.8-27B-GGUF Qwen3.8-27B-Q8_0.gguf --local-dir.
hf download harness-hub/Qwen3.8-27B-pi-adapted-GGUF Qwen3.8-27B-pi-lora-F16.gguf --local-dir.

llama serve -m Qwen3.8-27B-Q8_0.gguf --lora Qwen3.8-27B-pi-lora-F16.gguf \
    --chat-template-kwargs '{"enable_thinking": false}' -c 65536 --port 9931 --alias qwen3.8-pi

Since the adapter was trained with BF16 weights, matching the base GGUF closer to BF16 yields behavior closer to training. The card recommends Q8_0 as the default, noting that lower-bit quantization may cause behavioral deviations. The GGUF version has not been re-evaluated.

In addition, the following methods are also introduced:

  • Merging into the base model using PEFT to run on engines without LoRA support
  • Running with transformers serve
  • Loading with macOS Llama.app

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-19 harness-hub GGUF harness-hub/Qwen3.8-27B-pi-adapted-GGUF F16 0.4GB (fits in 4GB VRAM)

File sizes of each build:

  • Available builds in harness-hub/Qwen3.8-27B-pi-adapted-GGUF: F16 0.3GB

In addition, 13 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-19: Added converted builds to “Quantized and Converted Variants”: harness-hub/Qwen3.8-27B-pi-adapted-GGUF
  • 2026-11-19: Updated the hardware requirements table with the actual file sizes of harness-hub/Qwen3.8-27B-pi-adapted-GGUF.
  • 2026-11-19: Changed the title to show what the article covers (VRAM requirements, file list, etc.).
  • 2026-11-20: Added our own measurements: Japanese token efficiency, what is inside the GGUF file.

This article contains unverified information. We will append an update note once it is confirmed by a primary source.