PerGB

Decision models on your GPU

A decision model doesn't write text: it reads the input once and scores the options you give it, like routing a request or picking an action. So its speed is the time per question, not tokens per second, and most fit on a cheap card.

The open models

ModelSizeMemorySpeed reportedReported scoreLicenseHow to run it
Julia-1Supersonic Labs0.144B0.7 GB3 ms on an RTX PRO 6000—Apache 2.0llama.cpp
Laya0.421B1.0 GB5 ms on an RTX PRO 6000—Apache 2.0llama.cpp
Kev-4B4B3.6 GB12 ms on an RTX PRO 6000—Apache 2.0llama.cpp
lev4B3.6 GB36 ms on an RTX PRO 6000—Apache 2.0llama.cpp
OpenJev27B17.4 GB43 ms on an RTX PRO 6000—CC BY-NC 4.0: not for commercial usellama.cpp
ClefCloudflare27B17.4 GB209 ms on an H200 (Cloudflare's own test, as a service); estimated here from OpenJev, the same size—Apache 2.0llama.cpp, Workers AI
Perplexity Decider v1.2Perplexity27B17.4 GBNot published; estimated here from OpenJev, the same size69.42, Jev Decision IndexApache 2.0 (reported; one source says the repo is private)Its own code
mammoth-1Base Labs27B18.6 GBNot published; estimated here from OpenJev, the same size71.24 to 75.34, Jev Decision Index (sources differ)MITThe open weights

Memory: at Q4 with an 8K context; estimated from the model's size, which matches the files published. Speed: the median per question as its makers report it, on inputs that differ, so a number from one test isn't comparable with another. Scores: as reported; benchmarks' versions differ. Check a model's license before using it in a product. Details checked November 2026.

The cheapest card that holds each one

Live on eBay
ModelCardPrice nowPer question (est.)
Julia-1GeForce RTX 2060 Super 6 GB$89.0917–37 ms
LayaGeForce RTX 2060 Super 6 GB$89.0929–62 ms
Kev-4BGeForce RTX 2060 Super 6 GB$89.0970–149 ms
levGeForce RTX 2060 Super 6 GB$89.09209–448 ms
OpenJevTesla P40$270.00223–478 ms
ClefTesla P40$270.00223–478 ms
Perplexity Decider v1.2Tesla P40$270.00223–478 ms
mammoth-1Tesla P40$270.00223–478 ms

One PCIe card, from the live listings, shipping included. Per question: scaled from the model's speed on an RTX PRO 6000 by the card's compute and memory bandwidth, within about half either way.

Compare two cards

On any GPU's page, "Compare with" shows both cards side by side, with decision models next to language models.