Interconnects: multi-teacher on-policy distillation is the 2026 frontier recipe

Nathan Lambert and Finbarr Timbers spend an episode of Interconnects on the question of what the actual post-training recipe looks like at frontier labs in mid-2026, and the pattern they keep landing on has a name: Multi-teacher On-Policy Distillation, or MOPD. Each lab trains several domain-specialist teachers separately, then has a general student sample its own trajectories. Each sample is supervised by the relevant teacher's distribution via reverse-KL, not by a static dataset. The student stays a generalist while inheriting expert behavior on each domain.

Timbers lists the models he believes follow this pattern: MiMo Flash v2 with six experts in January, DeepSeek V4 with more than ten in April, and Nemotron 3 Ultra in June with experts spanning reasoning, code, and math. Others, like MAI-Thinking-1 and GLM-5, stop at specialist RL without the distillation step. Kimi K2.5 takes a different fork by combining text and vision RL in a single loop. The episode treats MOPD as the 2026 update to the DeepSeek-R1 reasoning-first lineage that defined last year.

The most pointed line is not technical. Timbers says the complexity of the recipe pushed against the limits of AI2's organizational capacity. Post-training at the frontier is now as much an org chart problem as a research one: enough teachers to be experts, enough harness to run on-policy sampling at scale, enough taste to pick the mix.

Why it matters

If you train or fine-tune open models, the next step up is not a bigger learning rate or one more SFT pass. It is splitting the team into teachers and running the student against them on its own trajectories, which is harder logistically than algorithmically.