Reading a model's spec sheet · GenAI Field Guide

An annotated spec line. Each term tells you one thing about how the model behaves.

The vocabulary

Term	What it means	What it means for you
35B (parameters)	The model holds about 35 billion learned numbers. A rough capacity number, not a speed or quality ranking.	A small new model often beats a big old one. You only see this number for open-weight models.
Context window (8k, 128k, 1M tokens)	The most tokens the model can attend to in one call.	Shared with the preamble, tools, and files. See The context window.
Precision (16-bit, 8-bit, 4-bit)	How many bits store each parameter.	Lower precision means a smaller, faster model. The quality loss is small at 8-bit and can be real at 4-bit, especially on small models and on reasoning.
Quantization (Q4, Q8, "quant")	Converting a model to lower precision after training.	How a 35B model fits on one workstation. Hosted models handle this for you; you never choose it on a portal.
Open-weight (Llama, Mistral, Gemma, and others)	The parameters are published; anyone can download and run the model. The opposite is closed-weight (GPT, Gemini, Claude, Grok), reached only through the vendor's service.	Whether a model is hosted by a vendor or run on your own hardware is a separate question; an open-weight model can be either. Downloading weights authorizes nothing: only infrastructure your command has accredited may hold official material.
Version (5.x, 6, 3.1)	The generation.	Newer is usually better at reasoning; see Choosing a model.
Tier (Flash, Pro; mini; Haiku, Sonnet, Opus)	Size-and-speed variants of the same generation.	Fast tier for bulk and simple work; top tier for hard work. Watch the word "Pro": on one platform it is the bigger model, on another the deepest reasoning setting.
Reasoning / thinking	The model works through steps before answering, sometimes with a settable level.	Slower, better on hard problems. See Thinking modes.
Multimodal	Reads images, documents, audio, or video, not just text.	Why you can upload a scan or a video.
Knowledge cutoff	The date its training data ends.	Anything after that must come from search or your documents. Find the date in the vendor's model card or your portal's documentation; a model's answer about itself is a guess.
Benchmark (a percentage on a named test)	A score on a public exam, usually reported by the vendor.	A rough ranking, not your task, and often marketing until independently reproduced. Fit and verification matter more.