12B open Nemotron trained on Surge RL data beats GPT-5.6 on SWE-Bench Pro V2
echen · x · 2026-11-06
SurgeAI trained NVIDIA's Nemotron 3.5 Super VL (only 12B active parameters) on its RL data using Thinking Machines' Tinker. The model beat GPT-5.6 Terra's published score on SWE-Bench Pro V2 and improved across 11 other benchmarks: GDP.pdf +18.6pp, HANDBOOK.md +12.8pp, AutomationBench +12.7pp, Chartography +11.2pp, AdvancedIF +7.7pp.
Their claim: frontier performance doesn't have to come from frontier size — better data and post-training can get you there, and the surface has barely been scratched.
More from Models
- Artificial Analysis launches Intelligence Index v5 and new Progress Index, reveals eval roadmap — ArtificialAnlys · 2026-11-06
- GPT-6 Cyber 'lightyears ahead of everything', says researcher after OpenAI early-adopter reveal — Eric_Wallace_ · 2026-11-06
- Fastino open-sources GLiDE: five decision models (0.4B-27B) claiming a new Pareto frontier — HowDevelop · 2026-11-06
- V0.5 model released with technical report, Hugging Face weights and inference code — rohanpaul_ai · 2026-11-06
- Clairvoyance launches Local AI benchmark site: 31 models × 15 hardware configs — draginol · 2026-11-06
- webAI Claims 12B Model Beats Gemini and Claude, But the Benchmark Claims Raise Eyebrows — eurie_kim · 2026-11-06