NVIDIA's open midtraining recipe lifts SWE-bench Multilingual from 55.8 to 62.8 with 500B tokens
dair_ai · x · 2026-10-17
- The recipe: NVIDIA published an open midtraining stage inserted between pretraining and post-training, with SFT kept identical. It trains at 256K context on 500B tokens of natural software-engineering data — packed repos, pull requests, commit histories and build logs — plus a small share of agent traces.
- Results: On Nemotron 3.5 Lightning, SWE-bench Multilingual went 55.8→62.8, SWE-bench Verified 60.7→64.5, SWE-bench Pro 32.0→35.5; on the larger Nemotron 3.5 Super, DeepSWE rose 11.7→15.7.
- Evaluation: Base-model proxy metrics didn't predict post-training outcomes, so every data choice was scored with short fixed SFT runs. The recipe and ablations are public.
Related event: NVIDIA Open-Sources Midtraining Recipe to Boost Agentic Performance(2 posts)→
More from coding & agent
- Claude can now reverse-engineer and rebuild apps like TradingView and OBS without source code — PrajwalTomar_ · 2026-10-17
- One-prompt interactive pond website from Gemini shows single-file HTML chops — R3d-Gr33n-Blu3 · 2026-10-17
- 15 AI Tools Worth Knowing in 2026: From Firecrawl to Agent Zero — nikola_mr64990 · 2026-10-17
- Your agent eval gate passed, but the next served request still scored zero — papaparija25 · 2026-10-17
- clauden lets Claude Code run on NVIDIA's free hosted models, dev says nemotron nears Sonnet 5.5 — gmaggiotti · 2026-10-17
- Interactive 3D simulator walks you through the inside of an AI agent — SeveralSeat2176 · 2026-10-17