Model Gallery

Release timeline — model size over time

Each point is a model at its release date (X) vs total params (Y, log). Point labels mark activated params/token; colour = family. Models without a recorded release date are omitted.

Model size — total vs activated params

Each point is a model. Dense models have active = total (upper envelope); MoE models fall well below (only a fraction of experts fire per token). X is log, Y is linear.

Per-request KV cache — BF16

Total KV cache one request occupies vs context length, assuming a BF16 cache. Architecture only — independent of GPU/parallelism. MLA (DeepSeek) and sliding-window/linear backbones grow far slower than full GQA/MHA.

Reference deployment — drives the prefill & decode charts below
Filter models — narrows every chart (no recompute) · default 30B+ total; clear the min to see smaller models

Prefill TTFT — SOL, batch size 1

Time-to-first-token vs context length. One line per model.

Decode TPOT — SOL latency

Time-per-output-token vs context length, at fixed batch size (lower is better). One line per model.

batch size = 1
batch size = 64

Decode throughput — SOL, batch size = max

Tokens/s vs context length at each model's KV-bound max batch (higher is better). TPOT isn't comparable here — each model runs a different max batch — so throughput (= effective tokens/step ÷ TPOT) is the apples-to-apples decode metric. Gaps = no request fits at that context.