Train Your Own Jev-style Classifier for $17 with Together AI
At a Glance
| Item | Value |
|---|---|
| Publisher | Together AI |
| Published | 2026-09-23 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
On September 23, 2026, Together AI published a method for building and training a custom Jev-style classification model using Qwen3.5 4B as the base model. This process uses approximately 38,000 data samples, keeping training costs to around $17 and completing in about 25 minutes.
Announcement Details
Together AI presents a procedure for building models with performance similar to the fast, low-cost classification model “Jev" on Together’s serverless platform. Specifically, it includes the following:
- Base Model and Dataset: Qwen3.5 4B is used as the base model, extracting and normalizing a total of 37,840 samples (noted as 38,340 samples in the documentation) from multiple datasets hosted on Hugging Face (MultiNLI, BoolQ, Banking77, AG News, Sst-5, Programmatic policies, Routing, Research taxonomy).
- Training Process: Together AI’s Fine-tuning service is utilized. By using the provided Python script, everything from data uploading to running the training job can be automated. Training takes approximately 25 minutes.
- Deployment: Once training is complete, the created model can be deployed as a dedicated HTTP endpoint. Deployment can be executed by specifying hardware such as NVIDIA H100 80GB SXM.
- Model Behavior: The model is designed to receive state, question, and options in JSON format, and return scores, boolean values, or option labels. For example, it handles tasks such as classifying intent from customer support messages, categorizing news articles, and performing text sentiment analysis.
Background
Together AI points out that Jev is a model rapidly gaining attention in the AI field. Jev is known as a powerful classification model that is fast and has extremely low operating costs. The purpose of this announcement is to demonstrate a method whereby users themselves can build and deploy such Jev-like classification capabilities at low cost and with high efficiency.
Impact on Local LLM Users
The method introduced in this article provides useful information for engineers who want to self-host lightweight classification models specialized for specific tasks at a very low cost. The impacts on readers include the following:
- Low-Cost Customization: You can build a custom classification model optimized for specific use cases (customer support intent interpretation, news classification, sentiment analysis, etc.) with a low budget of about $17 and a short time of about 25 minutes.
- Delivery Format: The trained model is used via an HTTP server through Together AI’s dedicated endpoint. Since it is deployed as an API endpoint, integration into existing software is straightforward.
- Model Behavior and Control: The model is designed to handle JSON-format input and output, making it easy to handle programmatically. However, technical notes also indicate that when calling the API directly or using the Chat Playground, you must explicitly set
temperature=0,max_tokens=8, andchat_template_kwargs={"enable_thinking": false}. - Use of Existing Models: Even without performing training yourself, you can directly use
together/Tev1-4B-experimental, which is published on Together’s serverless platform.
Related Articles
- Together AI Announces Canary Rollouts for Zero-Downtime Updates
- Scaling Coding Agent Traffic with GLM-5.2 and Dedicated Inference
- Together AI Explains Open-Source AI Stack and MIGHT Stack
- Kev: Lightweight Local Decision-Making Model Based on Qwen3.5