Qwen3.8-27B

Updated
21.08.2026
Thinking
Tools
Vision
Reasoning
Code
Multilingual
Run

Qwen3.8-27B is a dense 27.8B multimodal model with a 256K context that fits on a 24 GB GPU at 4-bit. Run it locally in Atomic Chat.

At a glance

  • License: Apache 2.0
  • Parameters: 27.8B dense, 64 layers, hybrid attention
  • Context length: 262,144 tokens
  • Modalities: text, image and video input
  • Minimum hardware: 12 GB VRAM at AD-IQ2_S, 24 GB for AD-Q4_K_M

Overview

Qwen3.8-27B is a dense 27.8B multimodal model from Alibaba's Qwen team and the smallest of the three models in the Qwen 3.8 release. It is also the only one of the three that runs on an everyday machine: a desktop with a 24 GB GPU, or a MacBook with 32 GB of unified memory. Alibaba published the weights on August 14, 2026 under Apache 2.0.

SpecificationQwen3.8-27B
Total parameters27.8B
ArchitectureDense, hybrid attention (Gated DeltaNet + Gated Attention)
Layers64
Context window262,144 tokens
ModalitiesText, image and video input
ReasoningThinking on by default, effort levels selectable, can be switched off
Multi-Token PredictionSupported, head included in the checkpoint
Release dateAugust 14, 2026
LicenseApache 2.0

48 of the 64 layers use Gated DeltaNet, a form of attention whose memory use stays the same no matter how long the conversation gets. Only the other 16 layers keep a regular KV cache, the part that normally grows with every token. Long chats and big documents cost far less memory here than on a typical 27B model. The model is also trained for Multi-Token Prediction, so it can draft several tokens per forward pass, the same mechanism behind speculative decoding.

Qwen3.8-27B benchmarks

Alibaba's launch numbers, from the model card, compare the 27B with its predecessor Qwen3.6-27B, the Qwen3.7-Plus API model, Muse Glimmer 30B and Claude Opus 4.6 Max:

Qwen3.8-27BQwen3.6-27BQwen3.7-PlusMuse Glimmer 30BClaude Opus 4.6 Max
Terminal Bench 2.1
Agentic terminal-use tasks
73.063.464.051.778.2
SWE-bench Pro
Real-world software engineering, harder split
61.753.557.651.253.4
LiveCodeBench v6
Competitive coding
90.383.989.6-88.8
GPQA Diamond
Expert-level science questions
89.287.890.383.591.3
Humanity's Last Exam
Expert-level questions across every domain
30.824.034.722.040.0
OSWorld-Verified
Computer-use agent tasks
84.363.973.365.972.7
SWE-MM
Multimodal software engineering
38.625.730.0-27.1

The 27B comes out ahead of Claude Opus 4.6 Max on four of the seven benchmarks. An Opus-level model now runs on a single 24 GB card.

Qwen3.8-27B hardware requirements

The system requirement to check is memory. We quantized the model from the original weights and published the builds as AtomicChat/Qwen3.8-27B-GGUF.

MemoryBuild to pickFile size
12 GBAD-IQ2_S11.1 GB
16 GBAD-IQ3_S13.8 GB
24 GBAD-Q4_K_M17.1 GB
32 GBAD-Q5_K_M20.2 GB
48 GB and upAD-Q6_K25.0 GB

Neighbouring files differ by a gigabyte or two, so when two builds both fit, take the larger one. That matters most below 14 GB, where quality falls fastest. If the GGUF format is new to you, start with what GGUF is, and see the best local LLMs for a 16 GB Mac for what else fits alongside it.

How to run Qwen3.8-27B in Atomic Chat

Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.

  1. Download Atomic Chat for your platform and open it.
  2. Search for Qwen3.8-27B in the model browser and open Download Options.
  3. Pick the build that fits the memory you have, then start a chat.

The full walkthrough, with the quant table and the llama.cpp commands, is in our guide to running Qwen 3.8 27B locally.

For the rest of the lineup, see every Qwen model you can run locally, or the trillion-parameter sibling Qwen3.8-2.4T-A95B.

License

Qwen3.8-27B is released under Apache 2.0. That permits commercial use, modification, and redistribution with no royalties, so you can build products on top of the model and run it on your own hardware without a usage fee.

Get the weights from Hugging Face

huggingface-cli download Qwen/Qwen3.8-27B
# 4-bit GGUF build from our own quant run:
huggingface-cli download AtomicChat/Qwen3.8-27B-GGUF Qwen3.8-27B-AD-Q4_K_M.gguf
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3.8-27B",
    "messages": [{"role": "user", "content": "Refactor this function and explain the change."}],
    "temperature": 0.6,
    "top_p": 0.95
  }'
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen/Qwen3.8-27B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Summarise this repository layout."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=1024)[0]))
import OpenAI from "openai";

const client = new OpenAI({ baseURL: "http://localhost:8000/v1", apiKey: "local" });
const res = await client.chat.completions.create({
  model: "Qwen3.8-27B",
  messages: [{ role: "user", content: "Write a Python function that parses a GGUF header." }],
});
console.log(res.choices[0].message.content);
Desktop
macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download

Frequently asked questions

Qwen3.8-27B is a dense 27.8B-parameter multimodal model from Alibaba's Qwen team, released on August 14, 2026. It reads text, images and video, carries a 262,144-token context window, and ships under Apache 2.0, so it can run locally in Atomic Chat without an API key.

Our AD-Q4_K_M build is 17.1 GB, so a 24 GB card such as an RTX 4090 runs it comfortably. On 16 GB, AD-IQ3_S fits at 13.8 GB, and 12 GB machines can still run AD-IQ2_S at 11.1 GB. A Mac needs roughly the same amount of unified memory.

Yes. The weights are published under Apache 2.0, which allows commercial use with no royalties or usage fees. You download them once and run them on your own hardware.

Yes. Once the weights are on disk the model runs entirely on your machine. In Atomic Chat your prompts, code and files never leave the device, which suits private or air-gapped work.

Qwen 3.8 Max is the 2.4-trillion-parameter MoE flagship, served through the Alibaba Cloud API and downloadable as Qwen3.8-2.4T-A95B. It needs a datacenter cluster. The 27B is the dense vision-language model in the same release and the only one that runs on a single consumer GPU.