Day 99: Open model landscape: Llama, Qwen, Mistral, Gemma, Phi, DeepSeek
Choosing from the open-model buffet
The AI Stylist runs on open models you control, not a closed API. The landscape shifts monthly, but the families are stable: Llama (Meta), Qwen (Alibaba), Mistral, Gemma (Google), Phi (Microsoft), DeepSeek. They differ in size, capability, context window, tokenizer, and — crucially — license. Picking well means matching the model to your constraint: a 7–8B model on a rented 24GB card for real work, a small GGUF-quantized model on the CPU droplet for the demo path.
- Parameters vs capability: bigger generally = more capable but heavier; a well-chosen 7–8B often suffices for a domain-scoped task like styling.
- Instruction-tuned vs base: use the *instruct* variant — it follows prompts; the base model only continues text.
- Family strengths differ: some excel at reasoning, some at multilingual, some at code; test on *your* task, not benchmarks.
Don't trust leaderboards — test on your task
Benchmark rankings are a starting filter, not an answer. A model topping a reasoning leaderboard may underperform on 'recommend formal shirts under ₹2500' compared to a lower-ranked one. Your Day-126 golden-question set is the real benchmark — evaluate candidate models on *that*, and let measured performance on your actual task decide.
Key terms
- Open model
- A model with downloadable weights you can run and control yourself, rather than access via a closed API.
- Instruction-tuned model
- A base model further trained to follow instructions/prompts (the "instruct" variant), suitable for assistant tasks.
- Parameter count
- The number of learned weights (e.g. 7B, 70B); loosely correlates with capability and resource cost.
For the AI Stylist, why might a well-chosen 7–8B model suffice rather than a much larger one?
