Redis Creator: Chinese AI Models Aren't Just Distilling US APIs

Recently, antirez (Salvatore Sanfilippo, author of Redis) published some content, the core of which refutes a claim: "The reason Chinese models are so strong is mainly because they distilled American models through APIs."

He believes this statement does not hold water based on machine learning principles, and is even "marketing from US labs" or a misunderstanding of ML.

antirez's point is: some people simply attribute the progress of Chinese open-source/closed-source models to "secretly using GPT/Claude APIs to generate data for distillation," which does not hold up in a strict machine learning sense.

This is actually due to inconsistent definitions, which will be understood later.

In his view, API distillation is impossible in the strict sense (white-box/soft distillation), because:

• True distillation requires extensive access to the teacher's complete logits (probability distribution), Chain-of-Thought reasoning trajectories, internal representations, etc.
• Commercial APIs only return the final text, providing none of this internal information.
• He compares API distillation to "only seeing a few points on a complex surface and wanting to replicate the entire surface," which is mathematically close to science fiction.

So white-box distillation of closed-source projects is basically impossible, remember, it's white-box (soft), because white-box distillation requires the student to not only output the same thing, but also to make its internal "thinking process" and knowledge representation as close to the teacher as possible.

Secondly, black-box (hard distillation) can be done, but its effect is limited:

• Using APIs to generate data for SFT (like Alpaca/Vicuna) can indeed improve 'response style', fill narrow knowledge gaps, or make the model better at following specific formats.
• However, it cannot create frontier-level general capabilities. Truly powerful underlying capabilities come from massive pretraining (trillions of tokens of data + huge compute).
• He explicitly states that even if you have traces generated by the teacher, without internal information like CoT, it is just "tuning the style or filling very small knowledge gaps at best."

Unless you are just using it to game benchmarks and climb leaderboards, you can distill to make benchmark data look good, but it will still fail in real-world use.

Furthermore, even with full model access, distilling a frontier model is extremely difficult:

• Many frontier Chinese models are already open-source, but many teams, including European labs, still struggle to train models that match their level.
• This shows that distillation/replication is not "easy to copy just because you have the data."

So he believes the real gap between Chinese and American models comes more from a compute deficit (restrictions on access to computing power), rather than simple technical copying or distillation. He does not deny the actual capability gap of current Chinese models, but firmly opposes attributing this strength mainly to "distilling American models."

However, in practice, black-box distillation can now be scaled up and can effectively transfer some task capabilities, output styles, format adherence, and a certain degree of reasoning patterns, so it is actually used quite a lot, but it truly cannot replicate the teacher's capabilities.

Of course, there are also opponents who say "using raw LLM responses for distillation is completely feasible, requiring only API access," citing Alpaca and Vicuna as examples trained this way.

This method does not require the teacher to expose any internal logits or hidden states, only the final text responses generated by the teacher, but the effect is certainly not going to be great.

Speaking of distillation, part of what antirez wants to express is that the term "distillation" is currently being severely abused and misunderstood. What people colloquially call "distillation" often conflates two completely different techniques:

• Hard Distillation (Black-box): Only uses the discrete token sequences (responses) generated by the teacher to train the student, using standard cross-entropy loss, similar to the method used by Alpaca and Vicuna mentioned earlier.
• Soft Distillation (White-box): Requires the teacher's complete probability distribution (logits), using KL divergence to make the student mimic the teacher's soft labels. This requires white-box access or API exposure of logits, which mainstream commercial APIs generally do not provide.

In fact, Apple's recent release of Apple Foundation Models is a distillation-based refinement, but closer to hard distillation (not entirely), because Apple stated that in the post-training/refinement phase, they used outputs from Gemini frontier models for refinement, meaning they used outputs (responses) generated by the Gemini frontier model to refine/align their own model.

Also, a 2026 paper, "Memorization Dynamics in Knowledge Distillation for Language Models" (arxiv.org/pdf/2601.15394), clearly distinguishes between the two and mentions that hard distillation is currently a feasible and common method in black-box API scenarios, although it will inherit more teacher-specific memorized samples than soft distillation.

So, is hard distillation viable? It definitely is, but its controllability, cost, and effectiveness are much worse.

In an interview, prominent figure Yao Shunyu also mentioned the differences between "hard distillation/smart distillation" at the practical strategy level. From an engineering perspective:

• Crude and brute-force: Directly having Claude/GPT generate a massive amount of tokens, then shoving them all into your own model for forced training. The expert's assessment is that it is "not very ethical commercially, and quite foolish from a governance perspective," indicating the company "hasn't figured things out and has no direction," because there is no data filtering, no strategy, and the teacher is not treated as a tool but merely as a "data printer."

• Smart distillation: Treating the strong model as an auxiliary tool and evaluator, strategically and purposefully integrating it into your own training system, for example:

  • Filtering high-quality synthetic data, using the teacher as a reward model/verifier, multi-agent collaboration to generate data, mixing real data with synthetic data, iterative self-improvement, etc.

Simply put:

• Crude hard distillation = low-level repetition (just dump data)
• Smart distillation = high-level engineering (build a smart data flywheel + training system)

At the time, antirez's view that it was not possible was largely based on white-box/soft distillation. He believes the real progress of Chinese models mainly comes from their own compute investment, data engineering, and research work, rather than being able to easily "distill" frontier capabilities through APIs. Treating the latter as the main explanation violates basic machine learning principles and underestimates the difficulty of truly building a strong model.

So this is where the divergence in the discussion arises. Although both are called "distillation," white-box/soft distillation is considered by the technical community to be the true distillation, while black-box/hard distillation, which only uses the text outputs generated by the teacher for SFT, is called Response Distillation.

So distillation has different meanings in professional and public domains, which is the reason for the disagreement. The rough difference is:

• Soft Distillation (White-box)

  • Has the student imitate the teacher's complete probability distribution (logits), not just learn the final answer.
  • The teacher's output is not a hard label (one-hot), but a soft probability distribution after temperature scaling.
  • The student can learn "dark knowledge": why the teacher thinks the probability of A is 0.7, B is 0.2, and C is 0.1, rather than simply being told it is A.

Of course, strictly speaking, the "soft" in soft distillation mainly refers to the probability distribution, not a natural language thought process. The teacher is not really telling the student "because...", but telling the student: "Among all candidate tokens/classes, my probabilities for A, B, and C are respectively..."

• Hard Distillation (Black-box)

  • Only uses the discrete token sequences (final output text) generated by the teacher, training the student with standard cross-entropy.
  • It's equivalent to treating the teacher as a "master craftsman" generating pseudo-labels, and then the student learns from this data like normal SFT.

In plain terms:

• Soft Distillation (White-box): The teacher not only writes the answer but also articulates the thought process and confidence: "I calculated 17 with 85% probability, because..." The student learns deep thinking and dark knowledge, becoming smarter.
• Hard Distillation (Black-box): The teacher only writes the final answer on the blackboard: "10+7=17". The student repeatedly copies the answer, learning the final result and format, but doesn't know how the teacher calculated it.

That is, one copies the thought process, the other copies the answer.

But in daily life, people generally don't understand or distinguish that much. So anyway, when I talk about distillation, I don't really differentiate much either, which is why I often get criticized or snarked at. But that's not important; what matters is what you think.

I also agree that pure hard distillation cannot produce a DeepSeek. On this point, antirez is indisputably correct.