Rubicon-preview (30B-A3B)
Rubicon-preview — a 30B-A3B model trained with a novel reinforcement learning framework built on rubric anchors, targeting open-ended, creative and humanities-centric tasks where verifiable rewards are hard to define.
Overview
Rubicon is a reinforcement learning framework that uses rubric anchors to train large language models for open-ended, subjective, and humanities-centric tasks — domains where verifiable rewards are hard to define. The released Rubicon-preview is a 30B-A3B parameter model built on the Qwen3-30B-A3B base.
Unlike conventional RLVR (Reinforcement Learning from Verifiable Rewards), which confines itself to code or math, Rubicon defines fine-grained rubric criteria as training anchors, enabling the model to be optimized for style, emotional expressiveness, and human-like writing.
30B total parameters with 3B activated per token (30B-A3B), trained under the Apache-2.0 license.
Highlights
- Token-efficient gains: +5.2% absolute improvement on subjective, humanities-centric tasks using only 5K training samples, outperforming DeepSeek-V3-671B by +2.42% on average.
- Stylistic controllability: rubric anchors precisely steer output style, producing responses that are more human-like, emotionally expressive, and less formulaic.
- Preserved general abilities: no degradation on general tasks — a common side effect of specialized RL — plus gains on reasoning benchmarks such as AIME 2024 (+4.1%).
Subjective & open-ended benchmarks (average of 7 tasks)
| Model | Avg |
|---|---|
| Qwen3-30B-A3B (base) | 65.29 |
| Rubicon-preview | 70.50 |
| DeepSeek-V3-671B | 68.08 |
