Blog · March 2, 2026
Qwen 3.5: The Small Models Revolution
A year ago, you needed a $10,000 GPU to run a decent vision-language model. Today, Qwen 3.5's 4B model runs on a MacBook Air and outperforms last year's 80B giants.
Why Qwen 3.5 matters
Alibaba's Qwen 3.5 family introduces a groundbreaking concept: small models with big brains. The 4-billion parameter model achieves performance on par with models 20× its size through a combination of:
- Native vision: Qwen 3.5 VL doesn't use OCR — it "sees" documents directly, understanding spatial layout, tables, and even handwriting.
- 1M token context: Process entire legal files (500+ pages) in a single pass.
- Mixture of Experts (MoE): The 35B-A3B variant activates only 3B parameters per token, making it faster than dense models 10× its size.
The hardware spectrum
What makes this revolutionary is the hardware range:
- 🔸 0.8B / 2B — Runs on a Raspberry Pi or smartphone. Perfect for edge IoT.
- ⭐ 4B — Runs on any modern laptop (16 GB RAM). Best price/performance ratio.
- 🔹 9B — Needs a GPU (6 GB+ VRAM). Maximum accuracy for complex documents.
- 🔷 35B-A3B (MoE) — For enterprise workloads on RTX 3090/4090 or Apple Silicon.
DataUnchain + Qwen = perfect match
DataUnchain uses Qwen 3.5 VL via Ollama to process documents locally. You choose the model size based on your hardware — from a Raspberry Pi in a warehouse to a workstation in an accounting firm.