What it is
A 2.8T-parameter open multimodal MoE with 104B active params, 1M context, and RL for coding, agents, reasoning, and vision. Instead of only scaling post-training on ~1T-class open models, it jointly scales pretraining and test-time RL: a 2.8T MoE with long-context KDA, depth-wise attention residuals, sparse 896-expert routing, then multi-effort agentic/coding RL distilled into one model.
Key results
- ProgramBench: 76.9, above GPT-5.6 Sol 72.9 and Fable 5 71.7
- AutomationBench: 30.8, above GPT-5.6 Sol 29.7 and Fable 5 29.1
- SWE-Marathon: 46.0, above Opus 4.8 41.0 and GPT-5.6 Sol 35.0
Shows open models can reach near-frontier agentic and coding performance by scaling both the base model and long-horizon RL together.