awakenedsniper78/ qwen3.8-turbo:latest

281 Downloads Updated 1 week ago

Qwen3.8-27B TURBO, uncensored, sized for 16 GB GPUs: 108k context, fast thinking, works in Claude Code and Codex.

ollama run awakenedsniper78/qwen3.8-turbo

curl http://localhost:11434/api/chat \
  -d '{
    "model": "awakenedsniper78/qwen3.8-turbo",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='awakenedsniper78/qwen3.8-turbo',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'awakenedsniper78/qwen3.8-turbo',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 week ago

1 week ago

2d7ef841fcae · 11GB

model

archqwen35

·

parameters27.3B

·

quantizationIQ3_XXS

11GB

params

{ "min_p": 0, "num_ctx": 110592, "presence_penalty": 0, "repeat_penalty": 1, "te

111B

Readme

Qwen3.8-27B TURBO: 16 GB GPU edition

A fast, uncensored Qwen3.8-27B that fits on a single 16 GB graphics card with 108k context, and still leaves room for your desktop and a browser. Tested on an RTX 5080.

Thinks less, answers sooner: a few hundred thinking tokens where stock Qwen3.8 often spends thousands
Uncensored: refusals removed by the original author (Heretic)
Works in Claude Code and OpenAI Codex: chat template fixed so agent tools don’t error out
Tool calling supported

Quick start

ollama run awakenedsniper78/qwen3.8-turbo

Required server settings

The 108k context only fits with a quantized KV cache. Set these for the Ollama server, then restart Ollama:

OLLAMA_FLASH_ATTENTION=1
OLLAMA_KV_CACHE_TYPE=q4_0
OLLAMA_GPU_OVERHEAD=536870912

Desktop already using a lot of VRAM? Lower the context inside a chat:

/set parameter num_ctx 65536

Performance (RTX 5080, Ollama 0.40)


Context	110,592 tokens
Speed	~58 tokens/s
VRAM	12.9 GB, 100% on GPU
Download	11 GB (3-bit IQ3_XXS)

Want ~100 tokens/s? The Hugging Face GGUF keeps the MTP head for speculative decoding in llama.cpp. Ollama doesn’t use it yet.

Recommended sampling

Temperature 0.6, top_p 0.95, top_k 20 (DavidAU’s advice for coding and tool calling).

What’s inside

Base: Qwen3.8-27B by the Qwen team
Fine-tune: DavidAU’s TURBO-Fable-Cold-Fusion-735-882 Heretic Uncensored
Quant: 3-bit IQ3_XXS using Unsloth’s dynamic per-tensor recipe and mradermacher’s importance matrix
Text only: no vision, so the VRAM goes to context instead

No coding benchmark has been run on this 3-bit version yet. A community tool-calling test on the 8-bit version put it slightly ahead of stock Qwen3.8-27B.

Credits

DavidAU for the fine-tune · Qwen team for the base model · Unsloth for the quant recipe · mradermacher for the imatrix · License: Apache 2.0

Full details and rebuild steps: Hugging Face repo

# Qwen3.8-27B TURBO: 16 GB GPU edition

A fast, uncensored Qwen3.8-27B that fits on a single 16 GB graphics card with **108k context**, and still leaves room for your desktop and a browser. Tested on an RTX 5080.

- **Thinks less, answers sooner:** a few hundred thinking tokens where stock Qwen3.8 often spends thousands
- **Uncensored:** refusals removed by the original author (Heretic)
- **Works in Claude Code and OpenAI Codex:** chat template fixed so agent tools don't error out
- **Tool calling** supported

## Quick start

```
ollama run awakenedsniper78/qwen3.8-turbo
```

## Required server settings

The 108k context only fits with a quantized KV cache. Set these for the Ollama server, then restart Ollama:

```
OLLAMA_FLASH_ATTENTION=1
OLLAMA_KV_CACHE_TYPE=q4_0
OLLAMA_GPU_OVERHEAD=536870912
```

Desktop already using a lot of VRAM? Lower the context inside a chat:

```
/set parameter num_ctx 65536
```

## Performance (RTX 5080, Ollama 0.40)

| | |
|---|---|
| Context | 110,592 tokens |
| Speed | ~58 tokens/s |
| VRAM | 12.9 GB, 100% on GPU |
| Download | 11 GB (3-bit IQ3_XXS) |

**Want ~100 tokens/s?** The [Hugging Face GGUF](https://huggingface.co/Jfish123/Qwen3.8-27B-TURBO-16GB-GGUF) keeps the MTP head for speculative decoding in llama.cpp. Ollama doesn't use it yet.

## Recommended sampling

Temperature 0.6, top_p 0.95, top_k 20 (DavidAU's advice for coding and tool calling).

## What's inside

- **Base:** Qwen3.8-27B by the Qwen team
- **Fine-tune:** DavidAU's TURBO-Fable-Cold-Fusion-735-882 Heretic Uncensored
- **Quant:** 3-bit IQ3_XXS using Unsloth's dynamic per-tensor recipe and mradermacher's importance matrix
- **Text only:** no vision, so the VRAM goes to context instead

No coding benchmark has been run on this 3-bit version yet. A community tool-calling test on the 8-bit version put it slightly ahead of stock Qwen3.8-27B.

## Credits

[DavidAU](https://huggingface.co/DavidAU) for the fine-tune · [Qwen team](https://huggingface.co/Qwen) for the base model · [Unsloth](https://huggingface.co/unsloth) for the quant recipe · [mradermacher](https://huggingface.co/mradermacher) for the imatrix · License: Apache 2.0

Full details and rebuild steps: [Hugging Face repo](https://huggingface.co/Jfish123/Qwen3.8-27B-TURBO-16GB-GGUF)

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