Distillation in 2026 (so far): which frontier models use it and how
This article is complementary material for Class 2: Distillation of the Training an Agent series Ben and I are doing, where we teach the post-training techniques behind a coding agent, step by step. Class 1 covered SFT on traces.
It also pairs with the short history of distillation we published before the class, if you want the background first.
Distillation is widely used in the post-training recipes of 2026's frontier models. The three stages we discussed in the live session (off-policy, on-policy and self-distillation) map directly onto how labs use it in real life.
A large teacher and a smaller student
The original use is still everywhere! Take a large, expensive teacher and train a smaller student to match it.
Gemma 3's tech report tells us that its post-training "relies on an improved version of knowledge distillation from a large IT teacher" (IT = instruction-tuned). The brand-new Gemma 4 tech report describes a similar post-training recipe, so we can infer that some distillation is also involved. DeepSeek-R1-Distill is also a case of this, and we already covered it in the brief history article: reasoning traces from R1 were distilled into compact Qwen and Llama students via plain fine-tuning (SFT) on the teacher's text, the sequence-level flavor.
These are the two flavors of the off-policy stage we covered in the class: match the teacher's next-token distribution (soft labels, white-box) or train directly on the teacher's generated text (hard labels, works black-box). R1-Distill is the second one. Same teacher-student idea, different signal.
Merge RL experts into one model
The newer use is different, and it is the one most frontier labs converged on this year. Getting a single model to be good at everything through RL turns out to be really complex, because the skills gained in one training stage tend to degrade during the next one. The workaround most labs landed on is to train a separate RL expert for each domain (one for math, one for code, one for agentic tasks) and then distill all of them into one student while it generates its own rollouts. This is on-policy distillation, where the student writes and the teachers grade every token.
Something interesting I noticed while reading these reports is that the teachers here are usually not bigger models. They are checkpoints of the same base, the same size as the student, each pushed further in a single domain with RL. What makes them good teachers is specialization rather than scale.
- DeepSeek-V4 is the cleanest description of the pipeline. Each domain gets its own expert (SFT, then GRPO), and afterwards "a single unified model is trained through on-policy distillation", with the student optimizing the reverse KL loss against the specialist teachers.
- The name for the multi-teacher form comes from MiMo-V2-Flash: MOPD, Multi-Teacher On-Policy Distillation, later studied in its own paper. Domain teachers provide a dense, token-level signal on whatever the student generates.
- GLM-5, the report behind the GLM-5.x family, applies it across training stages instead of across domains. After their sequential RL phases, a final distillation pass recovers the capability that degraded along the way, and the teacher is an earlier checkpoint of the same lineage. One step away from the model teaching itself.
- Nemotron 3 Ultra, NVIDIA's flagship, adopts the multi-teacher form at scale: more than ten specialized teachers, each with its own domain pipeline, give the student dense token-level guidance on its own rollouts.
- Qwen3 uses the same mechanic in the classic direction, a big teacher and small students that generate and align their logits to it. Their report puts the cost at roughly 1/10 the GPU hours of RL, with better results.
Every lab justifies this with the same argument. A teacher can give the student feedback on every single token it produces, while a reward in RL is one number for the whole attempt (in the class we explained this). So distillation converges much faster on the exact behavior the student needs to fix. Thinking Machines' write-up on on-policy distillation is the clearest practitioner version of this argument I have read, and it comes with numbers, matching their RL baseline at a fraction of the compute.
When the teacher is you
The third use drops the separate teacher entirely.
Cursor's Composer 2.5 trains with self-distillation. They inject a hint describing the desired behavior into the context, and the model with the hint becomes the teacher for the same model without it. A per-token KL pulls the unhinted policy toward its hint-conditioned self, so the model ends up producing the behavior without needing the hint at inference time. This is what we called the privileged teacher in the live session. Cursor's Sasha Rush explains it in detail in this clip from Dwarkesh Patel (Rafa Nadal shows up in the analogy, which alone makes it worth watching).
Thinking Machines shows another self-teacher, using the same on-policy recipe from the previous section with only the teacher swapped. After fine-tuning on new domain data, they distill from the pre-fine-tune checkpoint to restore the behavior that fine-tuning erased, while keeping the new knowledge. In class terms, this is the earlier teacher. This is their pitch for continual learning, keeping a deployed model learning new things without forgetting the old ones. Notice it is the same problem GLM-5 solves across RL stages, just at personal scale.
The teacher does not need to be bigger, just better in context. Sometimes that is the model itself.
The takeaway
So this is where distillation stands in 2026, so far. It compresses big models into small ones, it merges RL experts into a single model, and it lets a model learn from a better version of itself. Under the names, though, they are all variations of one teacher-student mechanic, open in TRL at a scale you can reproduce (everything from the class).
If you want to see how all of it actually works, go watch the class.
More classes are coming in the series. Follow Ben and me to catch the next one.
