The way frontier AI models get trained after pre-training has quietly undergone its biggest shift in years. Multi-teacher on-policy distillation (MOPD) is now the dominant post-training recipe at the frontier, showing up in technical reports from DeepSeek, NVIDIA, and others. If you run a service business that depends on AI capabilities, the models you’ll use next quarter were built with this approach.
Nathan Lambert and Finbarr Timbers broke this down in a new episode of the Interconnects podcast, walking through every major post-training recipe from 2022 to today. The takeaway is clear: the old single-pipeline approach (fine-tune, then do reinforcement learning) can’t keep up. The industry has moved on.
What happened
- MOPD is the new standard. Instead of running one giant reinforcement learning (RL) stage that tries to teach a model everything at once, labs now train multiple specialist “teacher” models (one for math, one for code, one for agentic tasks, etc.) and then distill all of them into a single “student” model.
- MiMo Flash v2 introduced it in January 2026. It was the first clean articulation of the approach: SFT first, then train about 6 domain-specialist teachers, then consolidate via MOPD into one model.
- DeepSeek V4 and Nemotron 3 Ultra scaled it up. DeepSeek V4 uses 10+ domain experts. Nemotron 3 Ultra runs multi-round MOPD across two iterations with more than 10 teachers.
- Not everyone has adopted it yet. Microsoft’s MAI-Thinking-1 uses a multi-stage RL approach closer to DeepSeek R1. Kimi K2.5 and GLM-5 also skip MOPD. But the direction of travel is obvious.
The numbers
- DeepSeek V4 uses 10+ domain-expert teachers in its MOPD pipeline.
- Nemotron 3 Ultra runs MOPD over 2 iterations with 10+ teachers spanning reasoning, code, math, and agentic domains.
- MiMo Flash v2 uses roughly 6 domain-specialist teachers.
- The post-training recipe has gone through 4 distinct eras in just 4 years: InstructGPT (2022), open RLHF recipes (2024), reasoning RL (2025), and MOPD (2026).
- Llama 3’s recipe ran rejection sampling and DPO over 6 rounds. That kind of monolithic multi-round approach is exactly what MOPD replaces.
5 things service business operators should know about MOPD
- Models are getting better at everything, not just one thing. The whole point of MOPD is that specialist teachers don’t trade capabilities off against each other. Your AI coding assistant doesn’t get dumber at writing because it also learned math. That means the next generation of general-purpose models will be noticeably more reliable across tasks.
- Specialist quality is going up fast. Each teacher model gets trained on a narrow domain with focused RL. That’s easier to optimize than a single model juggling everything. Expect sharper performance in code generation, math reasoning, and agentic tool use specifically.
- Open-source models will follow. Lambert and Timbers discussed how OLMO-3’s simpler recipe was already pushing against organizational limits. MOPD is more parallelizable across teams, which means open labs and even well-resourced companies can adopt it. The gap between open and closed models should keep shrinking.
- The cost of post-training is shifting. Instead of one expensive, conflict-prone RL run, you get many smaller, cheaper specialist runs. This is good news if you’re fine-tuning or customizing models. The tooling and techniques that power MOPD will eventually trickle down to your workflows.
- Agentic capabilities are a first-class training target now. Nemotron 3 Ultra explicitly includes “agentic domains” as a teacher category. This is a signal that the models shipping in late 2026 will be substantially better at multi-step, tool-using tasks like research, data entry, and workflow automation.
The hot take
MOPD won because it solves an organizational problem, not just a technical one. Training one model to do everything in a single RL run requires one team to coordinate every capability trade-off in real time. That doesn’t scale. MOPD lets you hand each domain to a separate team, train specialists in parallel, and merge at the end. The labs that adopted it first aren’t necessarily smarter. They just have better org charts. And that’s the real lesson: post-training is now as much a management challenge as a research challenge. If you can’t parallelize your training org, you can’t ship a frontier model.
The Agency OS play
You don’t need to train frontier models to benefit from this shift. But you should understand it, because it directly affects which models you pick and how you build on top of them.
This week, look at the AI workflows you’ve already deployed (or plan to deploy). Are you using a single general-purpose model for everything? Start testing the newest releases from labs using MOPD, like DeepSeek V4 or Nemotron 3 Ultra, against your current setup. These models are likely stronger across the board on multi-step tasks, especially anything involving code, math, or tool use. If you run a law firm, test document analysis and citation checking. If you’re in ecommerce, test product categorization and inventory queries. The performance gaps on domain-specific tasks will tell you whether it’s time to swap.
Longer term, pay attention to how MOPD-style thinking applies to your own fine-tuning or prompt engineering. If you’re customizing a model for multiple use cases, consider whether training (or even prompting) separate specialists and routing between them might beat a single generalist setup. The frontier labs just proved that specialists-then-merge beats monolithic training. The same logic applies at smaller scales. Build narrow, then combine.