The machine was already on. Routing is the rest.

[2026-03-22 — BENCH]

Eight gigabytes is the whole argument

The published Vulkan bench on an AMD Radeon Pro WX 7100 records 8 192 MB total and 2 974 MB already gone before any specialist loads — compositor and resident services, not a footnote.

Plan inference from what is left. Spec-sheet VRAM is a marketing number; the JSON is the instrument.

[2026-03-22 — CLASSIFIER]

Classify first, spend second

SmolLM2-1.7B Q8 sits hot at ~1.78 GB. Median classification is 90 ms (mean 110, min 78). That is the cost of knowing which specialist is worth a cold load.

On this card, an 8B reasoner crawls at 1.8 tok/s. The 1.2B Q8 specialists hold ~102 tok/s. The routing decision is the product.

[FIELD NOTE — MARGINAL COST]

The real comparison isn’t cost per query — it’s whether the machine is already on. Local Model Router for Linux · public README

Against that bench, hosted GPT-4o is tabled at about $10 per 1K queries. The router’s line is $0.00 — electricity the workstation was drawing anyway.

[2026-04 — MODELCASCADE]

Three tiers, one fail-cheap path

ModelCascade ships LOCAL (Ollama / llama.cpp / vLLM at $0), FAST (hosted small), CAPABLE (frontier). Coverage as published: 74% / +18% / +8%.

Those percentages are the product’s stated split, not a new measurement from this sitting. The cascade is the schematic; the bench JSON remains the metal.

[LICENSE — MIT]

Your keys. No telemetry line.

The public contract is BYOK and self-host. Paid tiers, if any, add operational tooling — they do not rent the routing decision. Core is MIT.

Data and models are commodity. Routing is intelligence. waynecolt.com · live masthead

[THROUGHPUT — 1.2B Q8]

Polaris has a sweet spot, and it is small

Code specialist TERMINAL-LFM2.5-1.2B Q8: 102.0 tok/s, 12 s cold load. General LFM2.5-1.2B Q8: 101.7 tok/s. The 8B Q4 reasoner is documented as impractical on GCN at 1.8 tok/s.

That is why the cascade exists as a product rather than a bigger local model. Hardware refused the heroic path; the schematic accepted it.

[EQUATION — LIVE SITE]

Intelligence as routing over dependency

waynecolt.com publishes INTELLIGENCE(t) = ROUTING(t) × (1+λ)ⁿ / DEPENDENCY(t). The move is to raise routing quality and session count, not to add another vendor in the denominator.

Compound learning — shadow pairs, distillation, graph memory — is listed as research on the homepage, not as a finished meter in this file.

[OPEN — OPERATOR]

What this deck will not do

It will not backfill a year-over-year percentage that was not measured. It will not treat a homepage round-number as a lab result. Where a figure is published product copy, the card says so.

The arbitrage was always going to close. Route responsibly. ModelCascade · closing line