Receive updates from Andrejusb for free, starting right now.
We can deliver them by email, via your phone or you can read them from a personalised news page on follow.it.
This way you won't miss any new article from Andrejusb. Unsubscribe at any time.
Site title: Andrej Baranovskij Blog
Is this your feed? Claim it!
I explain improvements I'm adding into Sparrow to achieve better accuracy for structured data. I'm using a method, where I run OCR step first, then construct advanced prompt with injected OCR data. This prompt is sent along with image to Vision LLM for structured data retrieval. All this happens as part of a single pipeline.
I was running detail tests to compare accuracy for the same models (Qwen3-VL and Mistral Small 3.2) running on Ollama and MLX-VLM (recent 0.3.7 version). MLX-VLM runs faster, but with lower accuracy. The same is valid across different models.
I'm comparing the Qwen3-VL 8B BF16 and Qwen3-VL 30B Q8 models for OCR and structured data extraction tasks. Based on my findings, the quantized 30B model runs faster and with better accuracy than the 8B BF16 model, despite using more memory.
I run couple of tests with structured data extraction using newest Qwen3-VL model on Mac Mini M4 Pro with 64GB. I discovered the same Qwen3-VL model with the same level of quantantization performs differently on Ollama vs. MLX. It seems model conversion step is crucial and we must evaluate model performance on different platforms before going to production.
style="...
I run and compare newest Qwen3-VL models in Sparrow. Qwen3-VL models run fast and provide good accuracy.
