Open-weights decision model
Ask for private hosting, call it as an API, run it locally, or get it pre-installed.
Dedicated hosting · on request
Best for steady traffic and privacy.
Weights are not on Hugging Face as a ready checkpoint — request it and we package it. Tell us what you want to do and we look into it for you. We reply within one business day.
We’ll email you once. Nothing else.
Best for trying it or low volume.
On requestWant Bonsai-2-27B v2 PQ2_0 + Q8_0 MMProj as pay-per-token? Tell us. We add models based on demand.
We’ll email you once when it’s live. Nothing else.
Best for testing offline.
One installer checks your machine (or a cloud/SSH GPU node), downloads Bonsai-2-27B v2 PQ2_0 + Q8_0 MMProj and starts a local Jev-compatible (typed answers with probabilities) endpoint.
Best for data that must stay on site.
Ship Bonsai-2-27B v2 PQ2_0 + Q8_0 MMProj on a box you own. Tell us your setup and we quote within one business day.
Model facts
Benchmark data from Benchmark Heaven. Model names belong to their authors; we host the published weights.
Dedicated hosting — good to know
A private, OpenAI-compatible endpoint that serves only Bonsai-2-27B v2 PQ2_0 + Q8_0 MMProj on GPUs reserved for you, with API keys you control, a live status page, request counts and spend. The model is pinned to an exact published revision.
From your prepaid Decision Models balance, in blocks: one hour at a time for “always on”, 15 minutes at a time while warm for “scale to zero”. VAT is applied when you top up. The next block is charged just before the current one ends. When your balance can’t cover the next block, the endpoint stops — you are never billed beyond your balance. Optional daily and monthly caps.
After a quiet period (15 minutes by default) we stop the GPU. The next request wakes it; until it is ready the endpoint answers 503 with a Retry-After header. A cold start takes a few minutes: a fresh GPU host loads the server and the weights (the estimate for your choice is shown above). Great for batch jobs and prototypes; pick “always on” for live traffic.
EU — hosting in EU, built for GDPR, aligned with the EU AI Act. Runs on GPUs of an EU provider in EU data centres. Global — lowest price. GPUs from independent and data-centre providers worldwide. Choose EU for personal or regulated data.
No. Requests pass straight through to your model server; we keep only counts (requests, tokens, errors) for your dashboard. Your prompts are processed on the GPU that runs your endpoint — on Global that GPU belongs to an independent or data-centre provider, so choose EU for personal or regulated data.
