What we read on a model card before a PoC

2026-Q4

We pin an exact repo id (Qwen/Qwen3-8B, not “Qwen”) and open that card plus the LICENSE file in the same tab. Roundups and family blogs are how Gemma 3 vision ends up in a 270M demo. The gemma-3-270m-it card is text-only; images are documented for 4B and up. That mistake costs a day and a confused workshop.

Five fields, in this order

License on the weights. GitHub’s default LICENSE is often the training code. Gemma is Gemma Terms of Use. Llama 3.2 is the Community License. Qwen3 is Apache 2.0. DeepSeek Coder V2 Lite is MIT on the repo and a Model License on the weights. If legal will look at the PoC, this sentence is in the kickoff notes on day one. We do not “check licenses later.”

Intended use. FunctionGemma 270M is not a chat model. StarCoder2-7B is completion, not instruct. Wrapping either in a chat UI looks fine for five minutes and then fails “extract these fields as JSON.” If the paragraph says “not intended for dialogue,” we do not use it as the only model in the room.

Input and output for this size. Family cards lie by omission. Qwen2-Audio is audio in, text out. It will not speak. Qwen2.5-Omni will. We write in → out on the same line as the repo id so nobody “adds voice” in week three.

Chat template. This is where tool calling and “thinking” actually live. Mistral 7B Instruct v0.3 takes tools= in Transformers. Qwen3-8B documents enable_thinking on the card. Granite 3.3 8B uses thinking=True. SmolLM2 mentions function-calling training on the 1.7B instruct line, not on 360M. If the card is silent, we write not_stated and we do not put tools on the success criteria. Trying it in a scratch notebook is fine.

Parameter accounting, only if we might self-host. Dense widget ≈ GPU guess. MoE: total and active. Mixtral is 46.7B total, 12.9B active (post); the widget says 47B. Names also drift: R1-Distill-Qwen-7B reports 8B on the card. If we are on an API, we still glance at this so nobody orders a 47B box later. Details in Local inference vs an API.

We paste those five lines into the PoC folder. Related: How we score a 4-8 week PoC.