Qwen 3.5 vs Qwen3-VL: Is There a Qwen 3.5 VL, and Should You Upgrade?

By Abdullah Zulfiqar · Data checked 27 October 2026 · 9 min read

Qwen 3.5 vs Qwen3-VL: which is better for vision? Qwen 3.5. There is no separate “Qwen 3.5 VL”, because every Qwen 3.5 model reads images and video out of the box. On Qwen’s own model card, the Qwen3.5-397B flagship beats the Qwen3-VL-235B flagship on all 27 vision benchmarks we pulled, with big jumps in computer use (OSWorld 38.1 to 62.2), chart reading (CharXiv 66.1 to 80.8) and OCR (OCRBench 87.5 to 93.1). Even the small Qwen3.5-9B beats the three-times-larger Qwen3-VL-30B-A3B on 12 of 13 tests.

Verdict: If you’re starting a new project, use Qwen 3.5. It’s newer, does text and vision in one model, and scores higher at every size we could compare. Keep Qwen3-VL only if it’s already running well in production and you don’t have time to re-test, or if you need a size Qwen 3.5 doesn’t offer (like the 32B dense model).

How to read these numbers: all scores below come from Qwen’s official Qwen 3.5 model cards on Hugging Face, which list Qwen3-VL as a comparison. Both sides were tested by the same lab, which makes the comparison fairer than most, but they’re still vendor-reported. We haven’t run our own benchmarks on these models. Treat gaps of 1 to 2 points as noise.

Key specs side by side. Sources: Qwen3-VL GitHub and model cards, Qwen 3.5 model cards, Ollama.

Is there a Qwen 3.5 VL?

No. This is the most common point of confusion, so let’s clear it up first.

With Qwen 3 (2025), Alibaba shipped text models and vision models separately. The text models were plain Qwen3, and the vision versions were a separate line called Qwen3-VL (“VL” = vision-language). If you wanted a Qwen model that could read a screenshot, you had to pick a Qwen3-VL model.

With Qwen 3.5 (2026), that split is gone. Every Qwen 3.5 model, from 0.8B up to 397B, takes text, images and video in the same model. So when people search for “Qwen 3.5 VL”, the answer is just Qwen 3.5. The same is true for Qwen 3.6, which came out in April 2026.

Qwen3-VL arrived in stages through September and October 2025. Qwen 3.5 replaced it with one multimodal family in early 2026.

What is Qwen3-VL?

Qwen3-VL is Alibaba’s vision-language model series, released between 23 September and 21 October 2025. It came in six sizes: dense 2B, 4B, 8B and 32B models, plus two mixture-of-experts (MoE) models, 30B-A3B and the 235B-A22B flagship. Each size came in two editions: Instruct for fast answers and Thinking for step-by-step reasoning.

Its main features, from Qwen’s GitHub page:

• 256K context, expandable to 1M, so it can take in long documents and hours of video
• OCR in 32 languages, with better handling of blurry, tilted and low-light text
• Visual agent: it can work through computer and phone screens, finding and clicking buttons
• Visual coding: turning screenshots and mockups into HTML, CSS and JS
• Video timestamps: it can point to when something happens in a long video

Qwen3-VL-8B was especially popular for local use, since it fits on an 8GB GPU at 4-bit.

What is Qwen 3.5?

Qwen 3.5 is Alibaba’s 2026 model family. The 397B-A17B flagship went live on OpenRouter on 16 February 2026, and the small models (0.8B, 2B, 4B and 9B) followed in early March. The middle sizes are 27B (dense), 35B-A3B and 122B-A10B (both MoE).

The big change is that vision is built in. Qwen 3.5 was trained on text, images and video together, so one model handles chat, coding, agents and image understanding. It has a 262K context window (about 1M with YaRN), a thinking mode you can switch off, and supports 201 languages and dialects.

If you want to see how Qwen 3.5 compares with other open models, read our Kimi K2.5 vs Qwen 3.5 and Gemma 4 E4B vs Qwen 3.5 9B comparisons.

Benchmarks: Qwen 3.5 vs Qwen3-VL

Summary of the head-to-head results on Qwen’s own model cards.

Flagships: Qwen3.5-397B vs Qwen3-VL-235B

The Qwen3.5-397B-A17B model card lists Qwen3-VL-235B-A22B as one of its comparison models. Qwen 3.5 wins every row.

Selected benchmarks from the Qwen3.5-397B-A17B model card. Higher is better.

A fair caveat: Qwen3.5-397B is a bigger model (397B total parameters vs 235B). But it uses fewer active parameters per token (17B vs 22B), so it isn’t slower or more expensive to run per token. It’s simply a better model.

The ten biggest jumps. BabyVision is shown without the code tool so the comparison is fair.

The biggest gains are in computer use (OSWorld +24.1), medical images (MedXpertQA +22.4), basic visual reasoning (BabyVision +21.1), chart reading (CharXiv +14.7) and math from images (MathVision +14.0). The smallest gains are in plain OCR (CC-OCR +0.5) and phone agents (AndroidWorld +3.1), where Qwen3-VL was already strong.

Small models: Qwen3.5-9B vs Qwen3-VL-30B-A3B

This one is more surprising. The Qwen3.5-9B model card compares it against Qwen3-VL-30B-A3B, a model with more than three times the total parameters. The 9B wins 12 of 13 tests.

From the Qwen3.5-9B model card. The only loss is SimpleVQA.

The one loss, SimpleVQA, tests simple factual questions about images (“what brand is this?”). That kind of world knowledge scales with total model size, so it makes sense that the larger 30B model keeps a small edge there.

What about Qwen3-VL-8B, the most popular local size? Its model card only shows scores inside images, not as text, and the Qwen 3.5 cards don’t compare against it directly. We didn’t want to guess. But since Qwen3.5-9B beats the bigger Qwen3-VL-30B-A3B on almost everything, it’s very likely to beat Qwen3-VL-8B too.

Where Qwen3-VL still makes sense

Qwen 3.5 is better on paper, but there are a few honest reasons to stay on Qwen3-VL for now:

• It already works for you. If you’ve tuned prompts, parsing and output formats around Qwen3-VL, switching means re-testing. Don’t swap a working production model without checking your own data first.
• You want the 32B dense size. Qwen 3.5 has 27B dense, but no 32B. If your hardware and pipeline are built around 32B, Qwen3-VL-32B is still a solid model.
• You want a separate Instruct model with no thinking. Qwen 3.5 lets you switch thinking off, but Qwen3-VL Instruct was built for fast, direct answers from the start. Some OCR and captioning pipelines like that.
• Fine-tunes and tools. Qwen3-VL has been out longer, so there are more community fine-tunes and guides for it.

For anything new, though, start with Qwen 3.5.

Model sizes and running them locally

Ollama download sizes. Your total memory need is a bit higher once you add context.

Both families are on Ollama, and both list text and image input there.

To try them:

# Qwen3-VL
ollama run qwen3-vl:8b

# Qwen 3.5 (vision is built in)
ollama run qwen3.5:9b

Then drag an image into the Ollama app, or pass an image path in your prompt. For the full setup, see our run LLMs locally guide, and Ollama troubleshooting if something goes wrong.

Which Qwen 3.5 replaces your Qwen3-VL?

If you’re upgrading, match by memory. This map pairs each Qwen3-VL size with the Qwen 3.5 model that needs about the same memory.

Matched by Ollama download size. The 397B figure is Unsloth’s 4-bit memory estimate.

• Qwen3-VL 2B → Qwen3.5 2B. A little bigger, but still tiny. For very low-end devices, Qwen3.5-0.8B is even smaller.
• Qwen3-VL 4B → Qwen3.5 4B. About the same size.
• Qwen3-VL 8B → Qwen3.5 9B. The most common upgrade. Still fits an 8GB GPU at 4-bit, and beats even Qwen3-VL-30B-A3B on most vision tests.
• Qwen3-VL 32B → Qwen3.5 27B. Slightly smaller download. We don’t have a direct benchmark comparison for this pair, so test it on your own images.
• Qwen3-VL 30B-A3B → Qwen3.5 35B-A3B. Same 3B active parameters, so similar speed.
• Qwen3-VL 235B-A22B → Qwen3.5 122B-A10B or 397B-A17B. The 122B is much smaller (81GB on Ollama). The 397B is the one with the head-to-head wins above, but it needs a lot more memory.

Also worth knowing: Qwen 3.6 is out (April 2026), with 27B and 35B-A3B open models. If you’re moving anyway, check our GLM-5.2 vs Qwen 3.6 comparison too.

Which should you choose?

• Building something new with images, screenshots or documents: Qwen 3.5.
• Local vision model on an 8GB GPU or 16GB laptop: Qwen3.5-9B.
• Computer-use or browser agents: Qwen 3.5. The OSWorld jump (38.1 to 62.2 on the flagships) is the biggest gain in the whole comparison.
• OCR and document parsing: Qwen 3.5 is better, but the gap is smaller. If Qwen3-VL already reads your documents well, upgrading is optional.
• Existing Qwen3-VL pipeline that works: keep it, test Qwen 3.5 on your own data, and switch when you’re ready.

Frequently asked questions

Is there a Qwen 3.5 VL model?

No. Qwen 3.5 doesn’t have a separate VL line, because every Qwen 3.5 model can already read images and video. If you want the Qwen 3.5 version of Qwen3-VL, just use the normal Qwen 3.5 model in the size you need, such as Qwen3.5-9B for local use.

Is Qwen 3.5 better than Qwen3-VL?

Yes, on Qwen’s own benchmarks. Qwen3.5-397B beats Qwen3-VL-235B on every vision test on its model card, including OSWorld (62.2 vs 38.1) and OCRBench (93.1 vs 87.5). At the small end, Qwen3.5-9B beats Qwen3-VL-30B-A3B on 12 of 13 tests.

What is the Qwen 3.5 equivalent of Qwen3-VL-8B?

Qwen3.5-9B. It’s a similar size (6.6 to 7.6GB on Ollama vs 6.2GB for Qwen3-VL-8B), fits on an 8GB GPU at 4-bit, and handles both text and images. It even beats the larger Qwen3-VL-30B-A3B on most vision benchmarks.

Is Qwen3-VL still worth using?

Yes, if it already works well for you. It’s a strong model for OCR, video and screen understanding, and it has more community fine-tunes. For new projects, though, Qwen 3.5 is newer, scores higher, and does text and vision in one model.

Can Qwen 3.5 read images and video?

Yes. Every Qwen 3.5 model accepts text, images and video as input. It can read screenshots, charts, documents and photos, and answer questions about video. It doesn’t take audio input.

How much memory does Qwen3-VL need?

On Ollama, Qwen3-VL downloads are 1.9GB (2B), 3.3GB (4B), 6.2GB (8B), 20GB (30B-A3B), 21GB (32B) and 143GB (235B-A22B). Plan for a little more memory than the download size once you add context.

Are Qwen 3.5 and Qwen3-VL free for commercial use?

Yes. Both are released under the Apache 2.0 license, so you can use them in commercial products, as long as you follow the license terms.

Sources

Checked 27 October 2026:

• Qwen3-VL GitHub (release dates, features): https://github.com/QwenLM/Qwen3-VL
• Qwen3-VL-8B-Instruct model card: https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct
• Qwen3.5-397B-A17B model card (flagship benchmarks): https://huggingface.co/Qwen/Qwen3.5-397B-A17B
• Qwen3.5-9B model card (small-model benchmarks): https://huggingface.co/Qwen/Qwen3.5-9B
• OpenRouter, Qwen3.5-397B-A17B listing: https://openrouter.ai/qwen/qwen3.5-397b-a17b
• Ollama library: ollama.com/library/qwen3-vl and ollama.com/library/qwen3.5

You made it to the end!

Get the weekly briefing

New model prices, benchmark results and Claude Code tips. One email a week. Free.

Email address Subscribe free

No spam. Unsubscribe anytime. Privacy policy

Website

Keep exploring

ReviewWinston AI Detector Review 2026: Is It Accurate? We Ran 4 TestsBenchmarkSciCodeGuideHow to Delete Your ChatGPT Account (Web, iPhone and Android) in 2026