AI BriefingAI21 and Deel post-train Nemotron 3.5 Super VL to outperform Claude Opus 5.5 on global payroll at 50x lower cost
·2026.11.04 09:00
Key point
A three-week engagement used near-policy SFT and SDPO to lift the model's quality by 20.7 points over the base and 8 points over the frontier competitor.
Details
- AI21 and Deel post-trained NVIDIA Nemotron 3.5 Super VL to surpass Claude Opus 5.5 by 8 points on quality for global payroll while costing 50x less per 1M tokens.
- The model achieved a 20.7-point quality improvement over the base model, reaching 0.720 criteria met and 0.339 all criteria met on the dev set.
- Training used a sequential approach: near-policy SFT on 412 curated examples followed by SDPO (self-distillation policy optimization) on 698 questions.
- Near-policy SFT lifted performance from 0.513 to 0.657 criteria met by training on the model's own hint-guided rollouts to avoid catastrophic forgetting.
- SDPO improved results to 0.720 criteria met by using a privileged teacher to score the student's tokens, allowing learning from questions the model initially got wrong without requiring gold answers.
- The base model's primary failure mode was incomplete answers (hedging or missing figures) rather than factual errors, which the post-training process corrected by steering the model to use existing knowledge.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Figma's unreleased editing model, post-trained on Qwen 3.8 27B, beats top third-party model on internal evalsSecond Look Research: Combining Fictional Stories and Difficult Advice SFT Reduces Agentic Misalignment in Qwen 2.5 to 4%
Y Combinator Labs and Olam Labs run multi-agent AI swarms in Minecraft and Mindustry