Surge RL Trains NVIDIA Nemotron 3.5 Super, Gains 4.5-18.6pp Across 12 Benchmarks
echen · x · 2026-11-06
Surge AI detailed how its research team used Thinking Machines' Tinker to train NVIDIA Nemotron 3.5 Super VL (120B total, 12B active) on four RL datasets spanning software engineering, instruction following, long-horizon enterprise workflows, and multimodal knowledge work.
Key results
- Improvements on all 12 benchmarks, ranging 4.5–18.6 percentage points
- Matched or beat much larger frontier models: exceeded GPT-5.6 Terra on SWE-Bench Pro V2, Claude Opus 4.7 on Terminal-Bench 2.1, and GLM-5.2 (744B) on AutomationBench
- AutomationBench end-to-end workflow completion nearly doubled (15%→28%); HANDBOOK.md requirement satisfaction rose 55%→68%
- Learned habits like checking evidence and testing code generalized to unseen settings
The team also praised Tinker for letting them run parallel RL experiments without managing GPUs. Conclusion: high-quality targeted RL data unlocks substantial headroom in open models.
Related event: Surge AI RL-Trains Nemotron 3.5 Super on Tinker, Nearing GPT-5.6(6 posts)→
More from coding & agent
- Google open-sources Agile Deliberation, an AI framework that turns fuzzy classification rules into explicit ones — aigclink · 2026-11-06
- AI uncovers 10-year-old bug where any text over 256 chars works as a valid root password — andersonbcdefg · 2026-11-06
- vLLM open-sources System Duo: a fast/slow semantic router that only calls deep reasoning when needed — vllm_project · 2026-11-06
- "Throw inference at problems until they go away" — a workflow take on scaling compute — rmcwhorter99 · 2026-11-06
- Claude Code bug leaks user email via User-Agent header, victim reports dozens of scam emails daily — niloofar_mire · 2026-11-06
- Nous Research ships Hermes Agent v0.23.0 with desktop-in-browser and unified gateway — NousResearch · 2026-11-06