Open-source Intern harness: Opus 5.5 hands off to Hy4 Flash, 92.4% at 1/3 the cost
m4rkGS · reddit · 2026-11-20
The nace team released Intern (github.com/nace-ai/intern), an open-source coding harness where every task runs through three cooperating models:
- Frontier: the smart, pricey model (Opus, Fable, Sol) that reads code and plans the hard parts;
- Efficient: a fast model (Hy4 Flash, DeepSeek Flash, Sonnet) that carries the rest, and can run fully local;
- Decision: a small model (Jev/Drex Max) between turns that picks handover moments, reranks searches, prunes context and filters tool output.
Key results (Terminal-Bench 2.1, 87 tasks): Opus 5.5 handing off to Hy4 Flash scored 92.4% at $0.127/task; Opus 5.5 alone scored 91–94% at $0.28–0.39 across Pi, Hermes and OpenCode — Hy4 Flash handled two-thirds of turns for 1% of the cost.
Local experiments: DeepSeek V4.1 Flash + local Qwen27B scored 65.3% at $0.092/task — cheaper than all 22 harnesses tested while beating 17 of them (Qwen27B alone: 40.3%). Single-vendor Sol→Luna matched Sol-alone's 44/62 passes at 39% less cost and 48% less than Codex CLI.
Install via uv tool install, one OpenRouter key covers all models; the efficient model can run locally via Ollama/LM Studio/Kapitsa. GGUF quants for Hy4 Flash are also available.
More from coding & agent
- Open-sourced Claude AI team: PM, CTO, UX roles that challenge your product ideas — marilynika · 2026-11-20
- Dev shares MCP lessons: human-only publishing enforced by omitting the tool entirely — mariobalca · 2026-11-20
- Gordon Wetzstein: coding agent in the loop boosts optical NN accuracy 35% to 96.8% — GordonWetzstein · 2026-11-20
- Optics as Code: Stanford agent tunes optical neural nets from 35% to 96.8% via hardware-in-loop experiments — GordonWetzstein · 2026-11-20
- Post-acquisition: Cognition's Poke agent gets its own computer and a chat interface — marvinvonhagen · 2026-11-20
- Browserbase uses agents to patch, build, and deploy 40M lines of Chromium code in under a day — blaizedsouza · 2026-11-20