Together AI Open-Sources tev1-4B-experimental: A $17 Jev-Style Classifier on Qwen3.5 4B
Together AI released tev1-4B-experimental, an open-source Jev-style classifier on Qwen3.5 4B fine-tuned for $17. Weights, data recipes, and serverless hosting a
On September 24, 2026, Together AI engineer Hassan El Mghari (@nutlope) announced the release of tev1-4B-experimental, a lightweight Jev-style classification model fine-tuned on top of Qwen3.5 4B for approximately $17 in compute cost. In addition to publishing the full repository at togethercomputer/tev1 on GitHub, Together AI has deployed the model directly to its serverless cloud infrastructure, offering inference at $0.042 per million input tokens with free output tokens.
Image source: Together AI
A $17 Training Recipe and Open-Source Fine-Tuning Pipeline
In modern AI agent architecture, routing simple decisions away from massive reasoning models and delegating them to lightweight, specialized classifiers is becoming standard practice. The release of tev1-4B-experimental provides a clear, cost-effective blueprint for developers looking to train and deploy custom task routers.
- Compact Foundation Base: Built on top of Qwen3.5 4B, the model delivers strong analytical judgment and intent classification without the latency and compute overhead of massive architectures.
- Extreme Cost Efficiency: Training achieved a practical custom classifier for approximately $17 in compute cost. This approach eliminates the need for expensive large-scale training while quickly producing an intent-specialized model.
- Fully Open Training Pipeline: Hosted at GitHub (togethercomputer/tev1), the release includes model weights, complete data preparation recipes, and step-by-step tutorials. Developers can inspect the codebase directly or point coding agents to the repository to fine-tune specialized classifiers on private datasets.
Serverless Deployment and $0.042 per Million Token Pricing
Beyond releasing model weights and data preparation recipes, Together AI has integrated the checkpoint into its cloud platform for immediate production evaluation.
- Immediate Serverless Availability: The model is hosted under the identifier
together/Tev1-4B-experimentalon Together AI's serverless endpoints, accessible via standard API keys. - Zero-Cost Output Token Structure: Input inference is priced at $0.042 per million tokens, while output tokens are free ($0 per million tokens). Because classification tasks focus on categorical decision routing rather than lengthy prose generation, this pricing model removes token accumulation overhead.
Experimental Classifier Limitations and Production Routing
This release reflects a growing trend across the open-source community toward specialized, task-routed decision models. On September 20, 2026, Jared Palmer released the Kev family of Qwen-based open decision models, demonstrating the practical advantages of small-footprint routers. While Together AI's release significantly lowers the technical and financial barriers to creating domain-specific classifiers, engineering teams should evaluate key production factors:
- Experimental Checkpoint Status: As designated by the
experimentallabel, this release represents an initial checkpoint. Production deployments should validate domain-specific classification precision and recall on target datasets. - Strictly Decision-Focused Interface: The model is tailored for intent detection and category mapping rather than conversational dialog or text authoring. System architectures must structure prompts for classification outputs without expecting generative prose.
By placing compact decision classifiers in front of primary LLM pipelines, development teams can filter routine requests, trim inference budgets, and accelerate execution velocity.