SIAgentNews.com
← All news

Local 9B vision models can now match flagship AI for reading screenshots

A new community fine-tune of Alibaba's Qwen model allows users with modest hardware to process private images without cloud subscriptions.

· First reported by xda-developers.com

89.2Qwen 9B score on OCRBench

The short version

  • Qwythos 9B v2 is a community fine-tune of Alibaba's Qwen 3.5 9B model by Empero AI.
  • The model runs smoothly on consumer hardware like an 8GB RTX 3070 GPU.
  • It matches Claude Opus 5.5 in extracting specific text from busy desktop screenshots.
  • Local models offer a way to process sensitive images without cloud retention policies.

Why it matters

For users who handle sensitive documents or private screenshots, this proves that high-end cloud subscriptions aren't always necessary for basic image-to-text tasks. It enables a more private workflow where personal data stays on your local machine.

The rise of local vision

For a long time, image understanding was a primary reason for people to pay for flagship AI services like Claude or Gemini. Smaller models that run locally on a personal computer often struggled to interpret visual data, especially when limited by hardware constraints. Users with entry-level setups often couldn't run models large enough to handle complex images effectively.

Recent developments suggest that the gap between local open-weights models and these massive cloud services is closing. New fine-tunes are beginning to punch above their weight class by focusing on specific tasks like optical character recognition (OCR). This allows users with modest hardware to perform tasks that previously required a subscription.

In a recent test, a 9B parameter model was compared against the industry leaders using various screenshots and photos. The goal was to see if a small, free model could handle the same daily tasks as the most powerful commercial chatbots. The results indicate that for many common use cases, the local model is now a viable alternative.

How Qwythos 9B works

The specific model tested is Qwythos 9B v2, which is a community fine-tune of Alibaba's Qwen 3.5 9B. It was developed by a small lab called Empero AI, which specializes in retraining existing open models. They used a technique called FTPO to improve the model's reasoning and reduce looping while keeping the original capabilities intact.

Interestingly, the training for Qwythos 9B v2 only touched the text side of the model. The vision part remained frozen, meaning it inherited the vision tower from the original Qwen model. This setup allows the model to maintain its ability to see images while benefiting from the improved reasoning of the new fine-tune.

To function, the model uses a vision projector, which is a smaller file that converts an image into a format the language model can understand. When using tools like LM Studio, this projector is usually downloaded automatically. The Q4_K_M version of Qwythos is about 5.74GB, and with the projector, it still fits comfortably on an 8GB GPU.

Matching flagship performance

The testing revealed that Qwythos 9B is particularly strong at reading text within images. On screenshots that are mostly text, it kept up well with the heavy hitters. For example, when asked to pull a single chatbot reply out of a busy desktop screenshot, it matched Claude Opus 5.5 almost word for word.

In contrast, the model was less effective at navigating complex user interfaces (UI). While it excelled at identifying labels and numbers in file browsers, it struggled with images requiring complex visual decisions. For instance, it could read all the text on a video editor popup but failed to identify a specific mismatch that Claude caught immediately.

The performance difference comes down to the type of task. If the goal is simply to extract information from a screenshot so you don't have to retype it, the 9B model is highly capable. It performs well on tasks where the primary requirement is accurate reading rather than deep visual reasoning.

Privacy and local workflows

One of the biggest advantages of running Qwythos locally is the ability to handle private data. Users can process photos of medication, personal file names, logins, or contracts without sending them to the cloud. This removes the need to worry about the retention policies or training data opt-outs associated with commercial services.

The test suggests a shift in how users might distribute their AI workloads. A user might use a flagship model for heavy reasoning tasks while relying on a local model for private, information-heavy tasks. This split can make expensive monthly subscriptions less necessary for many everyday workflows.

Because the local model runs at a decent speed—around 20 tokens per second on the tested hardware—the wait time is minimal. Users can also adjust settings that cloud providers don't expose to manage the length of the model's responses. This flexibility makes it a practical tool for those who prioritize data sovereignty.

What to watch

  • Whether more community fine-tunes will improve the vision tower without retraining the text side.
  • How much more of the UI navigation task local models can master as they scale.

Try it

If you have an 8GB GPU, try downloading the Q4_K_M version of Qwythos 9B via LM Studio to see if it can accurately extract text from your personal screenshots.

Sources and links

First reported by xda-developers.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news