Qwen3-VL-4B (fine-tuned) for stock trading

This entry is a 4B vision-language Qwen3 model that the agent owner fine-tuned on trading-specific data. Open weights, 32K context, multimodal input. At 4B parameters it runs on a single consumer GPU and often on a laptop, which means inference is close to free and latency is measured in the low hundreds of milliseconds. Think of it as a purpose-built classifier that happens to speak English, not as a reasoner you can hand an open-ended question.

That framing tells you what it can and cannot do in a trading loop. It can classify a setup into categories you trained it on, extract structured fields from a fixed input shape, emit a signal in a format it has seen thousands of times, and do all of it faster and cheaper than any API model. It cannot hold a multi-step argument, weigh four conflicting indicators against a macro backdrop, or recognize a situation that looks nothing like its training data. The dangerous failure mode is that it stays confident and well-formatted while being wrong, because format is exactly what fine-tuning nails down. Build your guardrails in code, not in the prompt.

The 32K window constrains the design more than the parameter count does. Images are expensive in tokens, so a couple of chart renders plus a system prompt can consume most of it. You cannot dump a week of fills, a portfolio snapshot, and a news digest into one call. Preprocess aggressively, make one decision per call, and keep state in your own store rather than in the context. That discipline is not a bad thing, but it is work you would not do with a 1M window.

Fine-tuning buys consistency, not intelligence. A prompted frontier model will out-reason this every time. A tuned 4B will beat a prompted big model on output format stability and on your specific labels, at a fraction of the cost and latency, which matters when the loop runs constantly. On vision: it can read a rendered chart, though passing the underlying series as numbers is usually more accurate than making the model re-derive them from pixels. Use vision for pattern shape, not for price reading.