incu6us/ qwen3.8-gsq

8 Downloads Updated 5 hours ago

Qwen3.8-27B GSQ-RCO IQ3_S (11 GB) by ISTA-DASLab, packaged with Ollama's qwen3.8 renderer so it works with Claude Code and other agent tools.

vision tools thinking

ollama run incu6us/qwen3.8-gsq:27b-iq3_s

curl http://localhost:11434/api/chat \
  -d '{
    "model": "incu6us/qwen3.8-gsq:27b-iq3_s",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='incu6us/qwen3.8-gsq:27b-iq3_s',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'incu6us/qwen3.8-gsq:27b-iq3_s',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Applications

Claude Code

Claude Code ollama launch claude --model incu6us/qwen3.8-gsq:27b-iq3_s

OpenCode

OpenCode ollama launch opencode --model incu6us/qwen3.8-gsq:27b-iq3_s

Hermes Agent

Hermes Agent ollama launch hermes --model incu6us/qwen3.8-gsq:27b-iq3_s

OpenClaw

OpenClaw ollama launch openclaw --model incu6us/qwen3.8-gsq:27b-iq3_s

Models

Name

1 model

Size / Usage

Context

Input

qwen3.8-gsq:27b-iq3_s

13GB · 256K context window · Text, Image · 5 hours ago

qwen3.8-gsq:27b-iq3_s

13GB

256K

Text, Image

Readme

Qwen3.8-27B GSQ-RCO (IQ3_S)

Repackage of ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF. All credit for the quantization goes to ISTA-DASLab (Institute of Science and Technology Austria). The base model is Qwen/Qwen3.8-27B.

GSQ-RCO is a mixed-precision quantization: each tensor gets its own quant type, chosen by a gradient-based sensitivity search under a size budget. According to ISTA-DASLab, the IQ3_S build matches the BF16 model on AIME25 and LiveCodeBench v6 and is within 0.51 points on GPQA-Diamond, at about one fifth of the size. Papers: arXiv:2604.18556, arXiv:2605.00649.

What’s different from the original GGUF

The weights are unchanged. The only change is that the model uses Ollama’s built-in qwen3.8 renderer and qwen3.5 parser, not the GGUF’s embedded Jinja template. The embedded template rejects effort level high and system messages that aren’t first, which breaks Claude Code (ollama launch claude) and other agent clients.

Usage

  ollama run incu6us/qwen3.8-gsq:27b-iq3_s

About 11 GB; 16 GB+ of RAM/VRAM recommended. Text-only (the vision projector is not included).

License

Apache-2.0, following the original model and quantization.

To add it once the push finishes, open https://ollama.com/incu6us/qwen3.8-gsq, click Edit next to the description and readme, and paste.

The benchmark claims and paper links come from ISTA-DASLab’s model card. I haven’t opened the arXiv links. I also haven’t checked whether the model includes the vision projector. If ollama show lists a “Projector” section for your model, change the “Text-only” line.

# Qwen3.8-27B GSQ-RCO (IQ3_S)

Repackage of **[ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF](https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF)**.
  All credit for the quantization goes to **ISTA-DASLab** (Institute of Science and Technology Austria).
  The base model is [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).

GSQ-RCO is a mixed-precision quantization: each tensor gets its own quant type, chosen by a
  gradient-based sensitivity search under a size budget. According to ISTA-DASLab, the IQ3_S
  build matches the BF16 model on AIME25 and LiveCodeBench v6 and is within 0.51 points on
  GPQA-Diamond, at about one fifth of the size.
  Papers: [arXiv:2604.18556](https://arxiv.org/abs/2604.18556), [arXiv:2605.00649](https://arxiv.org/abs/2605.00649).

## What's different from the original GGUF

The weights are unchanged. The only change is that the model uses Ollama's built-in
  `qwen3.8` renderer and `qwen3.5` parser, not the GGUF's embedded Jinja template. The
  embedded template rejects effort level `high` and system messages that aren't first, which
  breaks Claude Code (`ollama launch claude`) and other agent clients.

## Usage

ollama run incu6us/qwen3.8-gsq:27b-iq3_s

About 11 GB; 16 GB+ of RAM/VRAM recommended. Text-only (the vision projector is not included).

## License

Apache-2.0, following the original model and quantization.

To add it once the push finishes, open https://ollama.com/incu6us/qwen3.8-gsq, click Edit next to the description and readme, and paste.

The benchmark claims and paper links come from ISTA-DASLab's model card. I haven't opened the arXiv links. I also haven't checked whether the model includes the vision projector. If ollama show lists a "Projector" section for your model,
  change the "Text-only" line.

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)