Qwen3.5 is a family of open models built by Qwen, the AI team at Alibaba Cloud. We made smaller versions of four of them, 0.8B, 2B, 4B and 9B, using compression techniques that our Artemis platform found. Our default downloads are 32 to 38% of the size of the originals. They are text-only: the vision part of Qwen3.5 is left out.
They are free on Hugging Face and Ollama.
Who these files are for
A smaller file helps wherever a model has to run close to the work instead of in a large data centre.
- On a laptop or a small device. The download takes about a third of the space of the original.
- Where data cannot leave the building. Banks, clinics, legal teams and public bodies often cannot send documents to an outside AI service. A model on their own hardware keeps the documents where they are.
- Where the connection is poor. Field engineers, ships, factories and remote sites need tools that keep working offline.
Typical jobs: sorting and routing messages, pulling fields out of forms and invoices, short summaries, and asking other software to do something, such as looking up a customer record.
Which size to pick
| Model | Original download | Our default file |
|---|---|---|
| Qwen3.5 0.8B | 1.52 GB | 0.57 GB |
| Qwen3.5 2B | 3.78 GB | 1.39 GB |
| Qwen3.5 4B | 8.42 GB | 2.71 GB |
| Qwen3.5 9B | 17.92 GB | 5.66 GB |
Bigger models usually answer better and need more memory. This comparison did not measure memory use or speed, so try the smallest one that does the job on your own hardware, then move up if it falls short.
How close they stay to the original
We gave every file the same six test documents, which we did not use to make or choose any file, and measured how far its predictions drifted from the original Qwen model. Less drift means the smaller file behaves more like the original.
The chart shows every file we tested at each size, ours and the public ones from Unsloth and bartowski, so they can be compared directly.
Purple is TurinTech. The original Qwen model is the reference here, so it has no bar.
| File | Download size | Difference from the original |
|---|---|---|
| TurinTech Q4_K_M | 1.280 GB | 0.0163 |
| Unsloth Q4_K_M | 1.281 GB | 0.0289 |
| bartowski Q4_K_S | 1.328 GB | 0.0313 |
| Unsloth UD-Q4_K_XL | 1.340 GB | 0.0195 |
| Unsloth Q5_K_S | 1.385 GB | 0.0121 |
| TurinTech Q5_K_M | 1.395 GB | 0.0110 |
| bartowski Q4_K_M | 1.396 GB | 0.0254 |
It opens on the 2B files; pick another size, or switch to tool calls.
Behaving like the original is not the same as being right more often. On tool calls, the files are close to each other at every size, and differences of a few tasks out of 60 do not show that one file is better.
Where others are as good or better
- At 2B, Unsloth's slightly smaller UD-Q4_K_XL file was closer to the original on one test.
- At 9B, Unsloth's larger UD-Q4_K_XL file is closer to the original than our default.
- Our files score a little below the original on some tests, for example coding at 2B, maths at 0.8B and following instructions at 4B.
- At 9B, our default file predicts ordinary text slightly worse than Unsloth's Q4_K_M by one measure (perplexity), even though it stays closer to the original.
- We make no speed claim. In an earlier test on one consumer GPU (an RTX 3060), our larger 2B and 4B files ran about 3% slower than bartowski's Q4_K_M files.
Try one
With Ollama installed:
ollama run turintech/qwen3.5-2b-artemis
Swap 2b for 0.8b, 4b or 9b to try another size.
How we measured
The comparison in the chart is an independent check from 15 October 2026. We downloaded 32 files and checked each against its published SHA-256. Drift is the mean KL divergence against Unsloth's BF16 conversions of the original weights, averaged over six documents with two 256-token chunks each. Tool calls are 60 fresh BFCL tasks, run greedily with thinking off. 0.8B to 4B ran on an AMD Ryzen 9 9950X CPU and 9B on an NVIDIA GB10 with CUDA, both with llama.cpp b11617. The documents were not used to make or choose any file, but we cannot rule out that similar text was in Qwen's own training data. This is a small sample, so it shows a direction, not a general result.
Original sizes in the table are the BF16 conversions used in this check; the cards list the official conversion sizes, which can differ slightly (for 4B, 8.67 GB). Download sizes are file sizes, not memory in use, and some public files include layers that are never loaded. Ollama's default settings differ from these tests. Each model card (0.8B, 2B, 4B, 9B) has its own tables, test details and limits. The chart above shows every per-file result from this check; the cards use their own test panels.
These files are not made or endorsed by Qwen or Alibaba Cloud, and the credit for Qwen3.5 belongs to its authors. Teams use Artemis to find improvements in their own code and models. Explore Artemis.
Browse the model catalog for downloads and model cards, or explore more Artemis discoveries.
