How Surge RL pushed Nemotron 3.5 Super toward frontier-level performance
echen · x · 2026-11-06
Follow-up with Surge AI's full blog post on the Nemotron 3.5 Super RL training. Four RL datasets (software engineering, complex instruction-following, long-horizon enterprise workflows, multimodal knowledge work) each produced separate checkpoints.
- 4.5–18.6pp gains across 12 benchmarks; AutomationBench completion 15%→28%, HANDBOOK.md 55%→68%
- Matches GPT-5.6 Terra (SWE-Bench Pro V2), Claude Opus 4.7 (Terminal-Bench 2.1), GLM-5.2 744B (AutomationBench)
- Biggest gains on agentic tasks; habits generalized beyond training data
- Suggests substantial headroom from targeted, high-quality RL
Related event: Surge AI RL-Trains Nemotron 3.5 Super on Tinker, Nearing GPT-5.6(6 posts)→
More from Models
- FLUX 3 models beat FLUX.2 in all 19 Text-to-Image categories; Klein tops $9/1k price frontier — ArtificialAnlys · 2026-11-06
- Anthropic chief architect Sam McCandlish's first post hints at internal-acceleration report, possibly hitting 2x RSP threshold — inductionheads · 2026-11-06
- 6,662 decisions benchmarked: decision model Mammoth-1 averages 80.9%, beats best small LLM by ~3 points — pournasserian · 2026-11-06
- Hesamation: OpenAI Models Beat Anthropic's Only at Certain Kinds of Games — Hesamation · 2026-11-06
- Meta's new 'walrus' checkpoint reportedly matches crouton-alpha quality — ChrisGPT · 2026-11-06
- Quick test: Gemini 3.1 Pro knows the answer in its weights, Flash doesn't — giffmana · 2026-11-06