Together makes classification fine-tuning a $17 job

Together AI’s Tev1-4B-experimental is a Qwen3.5-4B-based classifier that fine-tunes on 37,840 labeled examples for about $17 and takes roughly 25 minutes to train, according to Together’s training guide. The guide’s nearby prose says 38,340 examples, so builders should treat the headline cost as an estimate rather than a contract. Teams with high-volume intent, policy, or routing decisions should benchmark a small typed model before paying frontier-token prices; Microsoft’s small-model routing playbook shows where this architecture fits.