Engineer autonomously trains a Jev-competitive model with an agent swarm for $3.1k in 20 hours

denisyarats · x · 2026-09-22

denisyarats shared AutoJev, a weekend project where a swarm of agents autonomously trained a Jev-like model using an internal system.
Setup: the swarm got a devbox with one H200; it mixes astra and fable agents, with sol and luna for synthetic data generation and filtering; ran for 20 hours.
Cost: $3.1k total ($1.9k agents, $1.2k data).
Results: built on qwen3.8-27b, initially worse than Jev but competitive/better after training, with 260k context and multimodal input; accuracy and calibration both improved.
Process: agents tried RL first; it didn't work, so they fell back to SFT on high-quality synthetic data.
Deployment: Jev-compatible API, playground, and inference; 120ms p95 on an H200.

More from coding & agent