Surge AI RL-Trains Nemotron 3.5 Super on Tinker, Nearing GPT-5.6

Surge AI published a blog post detailing how the team used Thinking Machines' Tinker platform to run reinforcement learning training on NVIDIA Nemotron 3.5 Super VL (120B total parameters, 12B active). The training drew on four of Surge's in-house RL datasets, targeting software engineering, complex instruction following, long-horizon enterprise workflows, and multimodal knowledge work respectively, each yielding its own checkpoint.

Confirmed

Scores improved across all 12 benchmarks, with gains ranging from 4.5 to 18.6 percentage points, bringing overall performance close to GPT-5.6
The four RL datasets cover software engineering (including SWE-Bench Pro-related evaluations), complex instruction following, long-horizon enterprise workflows, and multimodal knowledge work
Training was done on Thinking Machines' Tinker framework/platform, with each dataset producing an independent checkpoint

Why it matters

The author, echen, argues that people greatly underestimate the headroom of open-source models: a model with just 12B active parameters, when trained on high-quality RL data, can reach frontier-level performance of models roughly 20x its size
This case demonstrates how "high-quality RL data + a reusable training platform" can amplify the capability ceiling of open-source models, offering a practical path forward for mid- and small-scale models

2026-11-06 ~ 2026-11-06 · 6 related posts

Primary sources

How Surge RL pushed Nemotron 3.5 Super toward frontier-level performance — echen · 2026-11-06 04:24 (UTC+08:00)
Surge RL Trains NVIDIA Nemotron 3.5 Super, Gains 4.5-18.6pp Across 12 Benchmarks — echen · 2026-11-06 04:42 (UTC+08:00)
Surge RL-Tuned Nemotron 3.5 Super Beats Models Up to 20x Its Size — echen · 2026-11-06 04:42 (UTC+08:00)

12B open Nemotron trained on Surge RL data beats GPT-5.6 on SWE-Bench Pro V2 — echen · 2026-11-06
Nemotron 3.5 Super + Surge RL matches GPT-5.6 Terra across 12 benchmarks — echen · 2026-11-06
[source] How Surge RL pushed Nemotron 3.5 Super toward frontier-level performance — echen · 2026-11-06
Surge researcher praises Tinker: parallel RL experiments without GPU management — echen · 2026-11-06
[source] Surge RL-Tuned Nemotron 3.5 Super Beats Models Up to 20x Its Size — echen · 2026-11-06
[source] Surge RL Trains NVIDIA Nemotron 3.5 Super, Gains 4.5-18.6pp Across 12 Benchmarks — echen · 2026-11-06