# Qwen3.8-27B: what it is, is it free and can you run it?

Alibaba's Qwen3.8-27B is a free, open AI model that runs on a strong home PC. What it does, file sizes, how to install it and benchmarks.

URL: https://aitoolsthisweek.com/qwen3-8-27b
Published: 2026-11-15T15:28:41.209Z
Updated: 2026-11-15T15:28:41.209Z

**Qwen3.8-27B is a free, open AI model from Alibaba's Qwen team that can chat, write code, and understand images and video, and it's small enough to run on a powerful home computer.** The weights are free to download under the Apache 2.0 license, and Ollama offers a one-line install.

Three months after release it's still near the top of Hugging Face's trending list, with more than **20,000 likes** and about **6.5 million downloads** in the past month on its Hugging Face page. Here's why people keep grabbing it, and what you need to run it.

## What Qwen3.8-27B is

Alibaba released Qwen3.8-27B in mid-August 2026. Within two days it was one of the five most-liked models on Hugging Face, according to [Alibaba's own write-up](https://www.alibabacloud.com/blog/alibaba-unveils-qwen3-8-27b_603463). The company says it matches Qwen3.7-Plus, a mixture-of-experts model about ten times its size.

From the [official model card](https://huggingface.co/Qwen/Qwen3.8-27B):

| | Qwen3.8-27B |
|---|---|
| Size | 27 billion parameters (a "dense" model: all of it works on every word) |
| Input | Text, images and video |
| Context window | 262,144 tokens, extendable to about 1 million |
| Thinking | On by default, can be turned off; effort levels low, medium, xhigh |
| License | Apache 2.0 (free, commercial use allowed) |
| Hosted version | Qwen Cloud, listed as "coming soon" |

"Thinking" means the model works through a problem step by step before it answers. That makes it more accurate on hard tasks but slower. Qwen warns that turning effort down in long agent tasks doesn't always save time, because the model may make more mistakes and retry.

## Can you run it on your own computer?

Yes, if your computer is strong enough. Alibaba says it runs on consumer-grade hardware and, with quantization, even on a laptop. Quantization shrinks the model by storing its numbers with fewer bits, at some cost in quality.

File sizes from [Unsloth's GGUF versions](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF):

| Version | File size | Good for |
|---|---|---|
| UD-IQ1_S (1-bit) | 6.2 GB | Experiments only, lowest quality |
| UD-Q3_K_XL (3-bit) | 13.1 GB | Tighter memory budgets |
| UD-Q4_K_M (4-bit) | 16.6 GB | The usual balance of size and quality |
| Q8_0 (8-bit) | 29.0 GB | Near full quality, needs lots of memory |

As a rule of thumb, your graphics card's memory (or a Mac's unified memory) has to hold the whole file plus some extra room for the conversation. Image and video input needs a small extra file (the 0.9 GB "mmproj").

**The easy way:** install [Ollama](https://ollama.com/library/qwen3.8/tags), then run `ollama run qwen3.8:27b`. Ollama also lists Mac-optimized `27b-mlx` builds and an 8-bit `27b-q8_0` build. Developers can serve it with SGLang, vLLM or TokenSpeed, which Qwen recommends for heavy use.

New to local AI? Our guides to [on-device vs cloud AI](/on-device-vs-cloud-ai) and [AI models on your phone](/ai-models-on-your-phone) explain the trade-offs.

## How good is it?

Qwen's own benchmark table compares it with its predecessor, its bigger sibling and Anthropic's older Claude Opus 4.6:

| Test (what it measures) | Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus | Claude Opus 4.6 Max |
|---|---|---|---|---|
| SWE-bench Pro (fixing real code) | 61.7 | 53.5 | 57.6 | 53.4 |
| Terminal Bench 2.1 (coding agent) | 73.0 | 63.4 | 64.0 | 78.2 |
| OSWorld-Verified (using a computer) | 84.3 | 63.9 | 73.3 | 72.7 |
| GPQA Diamond (science questions) | 89.2 | 87.8 | 90.3 | 91.3 |

These are Qwen's numbers from its own test setup, not independent results. Still, the jump over Qwen3.6-27B is large, especially for computer-use and agent tasks.

## What it means for you

If you have a gaming PC with a big graphics card or a Mac with plenty of memory, Qwen3.8-27B is a strong way to run a capable AI model privately and for free, with no account and no data leaving your machine. If you don't, it's not something you'll use directly yet; free hosted chatbots are easier. For bigger open models that need a data center, see [Kimi K3.1 Pro](/kimi-k3-1-pro) and [Inkling 2 Small](/inkling-2-small).

## Frequently asked questions

### Is Qwen3.8-27B free?

Yes. The weights are free to download from Hugging Face under Apache 2.0, which also allows commercial use. Running it on your own computer costs nothing beyond your hardware and electricity. A hosted version on Qwen Cloud is listed as coming soon.

### Can Qwen3.8-27B run on a laptop?

Alibaba says that with quantization (compression) it can even run on a laptop. In practice you need a machine whose GPU or unified memory can hold the file you pick; the popular 4-bit Q4_K_M file from Unsloth is 16.6 GB.

### Is Qwen3.8-27B better than ChatGPT or Claude?

Not across the board. Qwen's own table shows it ahead of an older Claude model (Opus 4.6) on some coding and computer-use tests but behind it on others, such as Terminal Bench 2.1 and GPQA Diamond. It's a strong free option, not a clear winner over today's paid flagships.

## Sources

- [Qwen/Qwen3.8-27B model card (Hugging Face)](https://huggingface.co/Qwen/Qwen3.8-27B)
- [Alibaba Unveils Qwen3.8-27B (Alibaba Cloud Community / Alizila, Aug 17, 2026)](https://www.alibabacloud.com/blog/alibaba-unveils-qwen3-8-27b_603463)
- [unsloth/Qwen3.8-27B-GGUF file list (Hugging Face)](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF)
- [qwen3.8 model tags (Ollama library)](https://ollama.com/library/qwen3.8/tags)
