Logit to Token Sampler
Below are realistic logits for the next token after the prompt "France have the best". Temperature reshapes the distribution; top-p (nucleus) cuts the tail. Then we sample one token.
Selected token—
Entropy—
Tokens kept—
Token probabilities
Why 2D, not 3D: a probability distribution over tokens is one value per token. Bars encode it exactly; a 3D view would add perspective distortion without adding information.
Worked example (fact): softmax(logit/T). At T=1 the logit 8.1 for "CBs" gives about 47% probability here. At T=0.3 it exceeds 99% and sampling becomes near-greedy. At T=2 the tail tokens like "vibes" get real chances. Top-p=0.9 keeps only the smallest set of tokens whose probabilities sum to 0.9, then renormalizes.
Quantization Tradeoff Visualizer
Weights are numbers. Storing them in fewer bits shrinks the model and speeds up memory-bound inference, but rounds away detail. Slide the precision for a 7B-parameter model.
Est. memory—
Rel. speed—
Est. quality—
Estimate vs fact: memory = params × bits / 8 is arithmetic (fact, weights only; activations and KV cache add more). The speed and quality curves are illustrative estimates; real results depend on hardware, kernel support, and quantization method (e.g. GPTQ, AWQ, k-quants).
Size vs quality
Why these building blocks matter
Temperature and top-p are decoding-time dials: they never change the model, only how confident or adventurous its choices are. Quantization is a deployment dial: it changes what the model physically is in memory. Get both core positions right and everything else — prompting, tooling, product — works around them, the way a defense-first team frees its attack.