Kefuro-30B
One model for the whole ticket: intent, escalation and the reply, each with a confidence score. 30B-parameter mixture-of-experts, about 3B active per token, built on NVIDIA Nemotron 3.5 Lightning and post-trained on support outcomes.
Introducing Kefuro Service beyond chat
Kefuro in 32 seconds, with sound: built only for customer service, 90%+ lower cost, 20× faster decisions, open-source code, private deployment.
≈15.4B input + 0.77B output tokens / month
Not A Chatbot
General chat models are tuned to please whoever is typing. Support needs something else: the right answer, inside the store's policy, in the shopper's language, every time.
A Specialist Model
Kefuro learns from real support conversations and what happened next: resolved, escalated, refunded, or rated badly. Good outcomes teach it; bad ones become its negative examples.
Your Data Stays Yours
Kefuro runs on GPUs you control, its code is open, and customer conversations never pass through a third-party model API. Safety and legal cases always go to a person.
Cost: about $50,000 a month on GPT-5.4 list prices vs. about $2,100 a month self-hosted on two RTX PRO 6000 GPUs, at our volume. Speed: intent and escalation decisions take about 50 ms vs. a median of about 1 s for the hosted model; a full reply takes 0.2–0.5 s.
Safety, legal threats and "let me talk to a human" go straight to a person. No model decides those.
One model reads the ticket and returns the intent, whether to escalate, and how sure it is.
Every decision carries a confidence score. Above your threshold Kefuro acts; below it, the ticket goes to your team.
Replies are written from order data and the store's knowledge base, in the shopper's language.
One model for the whole ticket: intent, escalation and the reply, each with a confidence score. 30B-parameter mixture-of-experts, about 3B active per token, built on NVIDIA Nemotron 3.5 Lightning and post-trained on support outcomes.
Serving, training and evaluation code under Apache-2.0 on Hugging Face. Run Kefuro's stack on your own GPUs and train your own adapter on your own data.
huggingface.co/kefuro/Kefuro-30B →An open benchmark for ecommerce support: multilingual tickets, policy traps and must-escalate cases, scored on outcomes rather than style. Free to download and reproduce.
huggingface.co/datasets/kefuro/Kefuro-Bench →Built for Independent brands (DTC) · Marketplace sellers · Cross-border sellers · Omnichannel retailers
We start with ecommerce, where our own support data comes from. Other industries come later.
Languages English · Español · Français · Deutsch · Italiano · 日本語
Kefuro-Bench v0: 60 support cases in six languages, scored by fixed rules with no LLM judge. The metric names follow the ones Intercom reports for Fin, so the definitions are comparable; the numbers are ours, on our cases.
Configuration: Nemotron 3.5 Lightning reply model (not yet fine-tuned) with Kefuro's decision layer and hard rules. v0 cases are synthetic; a human-verified ecommerce set comes next.
Early-access partners get the first Kefuro checkpoints, a seat in the evaluation, and a say in which industries we cover next.
Prefer email? contact@daboss.ai
Kefuro is a model built only for ecommerce customer service, plus open-source code to run and train it and an open benchmark to measure it.
They are excellent general models, but ecommerce support is a narrow job with its own failure modes: invented tracking dates, refunds the store policy does not allow, the wrong language. A specialist trained on real outcomes makes fewer of those mistakes and costs far less to run.
For now, yes. Our own support data comes from ecommerce, so that is where Kefuro is trained and measured first. We will add other industries once ecommerce is solid.
On GPUs you control: ours, or yours as a private deployment. Customer conversations never pass through a third-party model API.
At our current volume, the same traffic costs about $50,000 a month on GPT-5.4 list prices and about $2,100 a month on two self-hosted RTX PRO 6000 GPUs. It is an estimate from our own measurements, not a promise for every business.
Not yet. Kefuro-30B is in training. Early-access partners get the first checkpoints and a seat in the evaluation.
The code is: serving, training and evaluation, under Apache-2.0 on Hugging Face, and Kefuro-Bench is public. The trained weights are not published; customers get them as a private deployment.
Kefuro is operated by DABOSSAI INC, a California corporation, 160 E Tasman Drive, Suite 125-126, San Jose, CA 95134. The team builds it in San Jose.