DeepSeek & Kimi Face Benchmark Gap & GPT-6 Distillation Allegations

Benchmark Gap Widens in Artificial Analysis v5

On October 27, Artificial Analysis released its Intelligence Index v5. Google's Gemini 4 Carbon leads with 59 points, followed by Claude Opus 5.5 (58) and GPT-6 Astra (57). Among Chinese models, Kimi K3.1 Pro scores 46 (13 points behind the leader), Qwen3.8 Max (0902) scores 34, and DeepSeek V4.1 Flash 1017 (max) scores 32 — gaps of 25 and 27 points respectively. The updated evaluation system makes the distance between Chinese models and top global closed-source models appear larger than before.

Allegations of a Smear Campaign Against Open Models

The benchmark shift sparked skepticism on X. User Da7em posted: "I think there's a quiet campaign to smear open models." Commenters linked the phenomenon to potential IPO plans and commercial competition. Some questioned the plausibility of the large score gap between GPT-6.1 Terra (51) and DeepSeek V4.1 Flash 1017 (32).

The Information Reports Distillation of GPT-6 Astra

On October 28, The Information published a report accusing Chinese AI labs of distilling OpenAI's GPT-6 Astra. The article cites:

Moonshot AI distilled Astra during the development of its new model, Kimi K3.1 Pro, which was released last week, according to three people with direct knowledge of the matter. Similarly, DeepSeek’s V4.1-Flash 1017, an Oct. 17 upgrade to its V4.1-Flash model that demonstrated substantial improvements from the original version released in September, also benefitted from distilling Astra, according to two other people with direct knowledge of that. Until recently, Anthropic’s Claude models were the biggest target of Chinese distillation efforts.

The report adds that DeepSeek CEO Liang Wenfeng initially resisted distillation for the new V4 generation but relented under competitive pressure after seeing domestic rivals like Z.ai gain traction with GLM models developed via distillation.

Distillation: From Accepted Practice to Contested Boundary

In early 2025, DeepSeek-R1 openly released multiple distilled models with "Distill" in their names, treating distillation as a legitimate technique to lower deployment barriers and democratize advanced reasoning. OpenAI and Anthropic themselves use distillation to create smaller, faster, cheaper models. However, as closed-source providers increasingly forbid distillation of their models, the industry's boundary between knowledge use and technology sharing grows ambiguous.

Anthropic's Own Data Practices Questioned

Litigation materials reveal Anthropic's "Project Panama": spending tens of millions to acquire rare books, destructively scanning them (cutting spines, destroying physical copies), with internal language about "destructively scan all the books in the world." This raises the question: when an AI company learns from humanity's collective knowledge, what constitutes fair use? And when others want to learn from that company's model capabilities, where should the line be drawn?

Fields Medalist Deng Yu Criticizes Double Standards

At an October 25 forum, Fields Medalist Deng Yu stated:

"For AI companies, you cannot generously distill the entire world's web literature and human knowledge, then claim no one wants to distill you — in fact nobody wants to distill you, they just want to understand the chain of thought — but you feel others want to distill you and start saying this is not allowed, that is not allowed."

Technical Achievements Beyond Benchmarks

Regardless of unverified allegations, Chinese models demonstrate concrete technical progress:

Kimi K3.1 Pro : 3 trillion parameters, 1 million token context window — the largest open-source model globally — continuously narrowing the gap with top closed-source models.

DeepSeek : Innovates at the infrastructure layer with high-performance operators and communication components (DeepGEMM, FlashMLA, DeepEP) and the recently open-sourced DeepInfer inference framework, improving inference speed, efficiency, and deployment cost — gains not directly reflected in benchmark scores.

Therefore, even if gaps remain in certain capability evaluations, dismissing or negating these models' R&D and underlying technical contributions based solely on unproven leaks is unwarranted.