AI Compute RadarMain brain
Low cost
Qwen3.8 27B
$0.15/M input tokens
The most-adopted model for this use case among those with a published price. We publish no quality benchmarks, so this reflects usage, not a capability ranking.
01 / Stack builder
What are you building?
Hybrid stack
Cloud budget: $50/mo buys about 333M input tokens/mo at Qwen3.8 27BHybrid stack. Main brain: Qwen3.8 27B. Fast + cheap: GLM-5.2. Local model: GLM-4.7-Flash. Runtime: llama.cpp. Cloud router: OpenRouter.
Fast + cheap
Low cost
GLM-5.2
$0.01/M input tokens
Cheaper than your main brain, for classification, routing and other high-volume calls that do not need the strongest model.
Local model
No per-token cost
GLM-4.7-Flash
Q4_K_M · ~18.9 GBMemory fit 73/100
Runs on hardware you already own, so everyday work costs nothing per token and never leaves your machine.
Runtime
No per-token cost
llama.cpp
GGUF weights run under llama.cpp without a packaged runtime.
Cloud router
No cost
OpenRouter
One API across providers, so swapping a model does not become a migration. The prices on this site come from its public catalogue.
Reading the stack builder
As of today's readingsOn November 2, 2026 the builder draws on 49 tracked models, 49 of them with a measured local file, 37 hardware profiles and 8 rental classes with a measured median. The same five answers return the same stack until one of those readings changes.
The builder asks five questions and returns a complete stack with a reason next to every slot: the model, the file to download, the runtime, the hardware to run it on, and the rental fallback when the hardware you have does not fit. It is assembled by rules, not by a language model, so the same five answers always give the same stack, and every slot can be traced back to a measured fact: a file size, a fit verdict, a rental median.
The rules are the site's own: the fit engine that rates every model against every machine decides the model slot, the memory table decides the file, the rental price board decides the fallback. What the builder does not know is what you like. It cannot judge whether a model writes well for your task, and it does not try; it narrows the field to what runs, and leaves the taste test to you.
Read a slot's reason before taking its content. Where the builder says a model fits comfortably it means the usable memory of the chosen machine at the chosen context, not the sticker; where it names a rental class it means the cheapest class whose median it has measured and that runs the model comfortably, not the cheapest card on earth.