model-scale

What does it cost to run this model yourself?

How the numbers are computed, and where they lie to you

Memory. Weights are total_params × bytes_per_param — for a mixture-of-experts model that means every expert, because they all have to be resident somewhere. The active-parameter count buys compute and latency, not storage. KV cache is 2 × layers × kv_heads × head_dim × bytes per token for MHA/GQA, or layers × latent_dim × bytes for MLA. Sliding-window layers hold only their window; linear-attention layers hold a fixed state that never grows with context.

Decode throughput is a memory-bandwidth roofline: per step the GPUs stream the resident backbone, whichever experts the batch routed to, and every sequence's KV cache. The expert term is 1 − (1 − k/E)^batch, the share of the expert pool a batch touches — which is why MoE sparsity stops paying at large batch.

Per-user speed converges. As batch grows, the cache forces more GPUs and those GPUs bring bandwidth, so both sides of step_time = bytes / (gpus × bandwidth) grow together. Per-user speed approaches (1+overhead) × bandwidth × efficiency ÷ usable_per_gpu — a property of the card, not the model. Bigger cards are worse: the extra capacity fills with cache that must be re-read every step.

Seats are not people. A seat runs around the clock; a person uses it for part of a day. Worker-equivalents convert as seats = workers × hours_per_person ÷ (24 × utilisation). The two divisors are different things: hours-per-person is how long someone actually drives a session, utilisation is how much of a cluster's round-the-clock capacity bursty demand lets you consume.

What it does not model. Interconnect is folded into one achieved-bandwidth factor rather than costed per collective, which flatters wide parallel groups. Parallelism is assumed to shard cleanly at any GPU count. Continuous batching, prefill/decode interference, prefix caching and speculative decoding are all absent — TTFT is computed on an otherwise-idle replica. There is no cloud rental, depreciation, or cost of capital.

Units. Decimal bytes throughout: 1 GB = 10⁹ bytes, so 2.4T parameters at FP32 reads as exactly 9.6 TB.

Every preset — model architectures and GPU specs alike — is an approximate starting point transcribed by hand, and GPU prices are volume-dependent guesses. Edit them. A model labelled hypothetical is not a real announced product.

No dependencies, no build step, no tracking. Everything runs in the page; the URL hash is the only place your configuration is stored.