Chinese team EvoMap accuses Silicon Valley star AI project of plagiarism, Hermes Agent denies it twice
36氪·商业情报局（第二季） · 2026-04-16 · 5 min
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Chinese team EvoMap accuses Hermes Agent of copying its EvoVer self-evolution engine; Nous Research denies it twice.
Use cases
- EvoVer's ten-step evolution loop maps to Hermes self-evolution module's ten-step process
- EvoMap alleges systematic replacement of 12 sets of core terms (Gene to Skill MD, etc.)
- Time gap from EvoVer core architecture to Hermes function was only 24 to 39 days
- Hermes Agent reached more than 85,000 GitHub stars via its self-evolving skill ecosystem
- EvoMap changed license from MIT to GPL3.0 and obfuscated its core module in protest
KPIs / results
Tools / build
- Hermes Agent
- EvoVer engine
- GEP protocol
- EvoMap
🌐 This transcript was automatically translated to English from the original.
This column is exclusively produced by 36g. The equal amount comes from Jiemian News. On April 15, Chinese AI team EvoMap publicly stated that the core self-purification capabilities of Silicon Valley's star AI project Hermes Agent were systematically replicating its EvoVer engine. The latter responded twice with an almost arrogant attitude, once again triggering heated discussions on the ethical bottom line and innovation fairness in the open source world. In a technical report released by EvoMap It uses DTable to disclose timestamps, circular code structures, architecture designs, numbers and systems, evolutionary cycles and other cross-verifiable evidence chains. It points directly to the high degree of isomorphism between the Hermes Agent self-evolution module and the EvoVer engine. The core of the dispute points to the origin and attribution of the self-evolution architecture. EvoMap gives a clear and verifiable timeline. On February 1, EvoVer open sourced and disclosed the GEP core design. On February 16, the team released an in-depth analysis of GEP. A complete ladder of ScanSelect, Mutate, Validate, Semitify main loop and three-level memory system. On February 25, Hermes Agent released V0.1.0, positioned as the initial basic version. On March 9, Hermes Agent created a self-evolving independent warehouse. On March 12, Hermes Agent V0.2.0 launched the skills ecosystem. The core self-evolution function was officially launched. The time difference from the core architecture to the corresponding function of Hermes was only 24 to 39 days. This time window has become the key basis for EvoMap's accusation. Both parties adopt a closed loop that automatically extracts reusable experience after the task is completed. Both parties set up periodic self-evaluation and reflection mechanisms. Both realize skill runtime discovery and on-demand loading. The most convincing evidence is that EvoVer's ten-step evolution main loop corresponds to the ten-step execution process of the Hermes self-evolution module. There are also systematic replacements of 12 sets of core terms, such as Gene corresponding to Skill MD Capsule corresponding to skill execution records, Solidify corresponding to Skill Manage Create, etc. In all seven public materials of Hermes, the EvoVer and GEP protocols, which are the first open source, have zero references, zero lines, and zero attribution. EvoMap stated that in the first two months, the team frankly threw the core GEP protocol, three-layer memory system, and periodic reflection loop to the community because we feel that in this era of rapid AI development, transparency and openness are the coolest gestures we can make. We did not hide it and put all our cards on the table. But reality has taught us a lesson. In the face of detailed accusations, News Research, the laboratory to which Hermes Agent belongs, expressed its position twice. The first response was issued by the official account, saying that the project warehouse had been created in July 2025. It claimed to be the pioneer of the underlying technology of Hyundai Agent Soft Price and asked Evo Map to delete the account. Later, it deleted the post and blocked the other member. The second response came from News co-founder Technium. He said that I have never heard of this person in my life, his project or anything he is doing. Today is the first time I have heard of it. Or after seeing this project, they claim without evidence that I plagiarized their work. This is a lie. Hermes Fong has two rebuttal logics. One is that the warehouse was created earlier than Evover, proving that the technology was developed first. The other is that the bottom layer relies on academic open source projects. The architecture design department has independently converged. Evo Map then pointed out that although the Hermes warehouse was established earlier, it has been private for a long time. The basic version was only made public at the end of February. The questioned self-confidence ability was not fully launched until mid-March. The so-called first-mover advantage cannot cover the controversial functions. JEP and GEP belong to optimization algorithms and evolutionary protocols respectively. They cannot explain the high degree of overlap of the upper architecture, loop process, and memory system. Public information shows that Evo Map is a Shenzhen startup company founded in 2025 with a team of less than 20 people. The founder Zhang Haoyang is a post-95s generation. He became China's smallest Unity developer at the age of 14. He started his business at the age of 17 and later served as the technical planner of Tencent's Peace Elite. In 2023, he chose to start a business again and founded AutoGame. It has received three rounds of financing totaling tens of millions. Hermes Agent is a phenomenal project under the Silicon Valley Lab News Research. It has gained more than 85,000 stars on Github with its self-evolving skill ecosystem and is regarded as a representative of the next generation of intelligent agents. It has wide influence in the developer community. The core of Hermes Agent’s self-evolving function is that it has built-in complete execution evaluation, creation and improvement, and a closed-loop learning cycle, turning it from a one-time execution tool of traditional AI into a dynamic agent that can independently accumulate experience, continuously self-optimize, and become smarter with use. This is the most essential difference between it and other AI assistants. Although Hermes Agent denies plagiarism, there has been a precedent for Silicon Valley teams to learn from China's AI open source results. In 2024, the Lama 3 Micro-Motai model released by the Stanford team was accused of plagiarizing the wall-facing smart mini CPM Lama 3 V 2.5. Beijing community comparison found that the architecture, configuration, and special recognition capabilities of the two were highly consistent. In the end, the author involved publicly apologized and removed the model. Earlier, Meta engineers were exposed. In-depth analysis of the Chinese team. The MoeAE architecture tried to be copied, but due to lack of technical details, the project failed. The common trajectory of such incidents is that a small Chinese team relied on its own innovation ability to launch open source results. The Silicon Valley team quickly followed up and promoted it to the world with stronger resources and voice. The original party found it difficult to obtain reasonable attribution and recognition. In the end, most ended up with an apology and removal from the market, with few substantive compensations and institutional corrections. Therefore, EvoMap said that it has changed the core module of EvoWord to obfuscate release, and changed the license from MIT to GPL30. From individual developers to small teams, from large domestic manufacturers to overseas giants, open source code is not respected and is becoming the norm. The AI era has lowered the development threshold to a minimum. Everyone can create, but the dignity and living space of original creators have been compressed to the extreme. We can only use our own actions to avoid the original to express our attitude.