NetEase Yuyan 2

• 1. overview (674 chars)

  "NetEase Yuyan 2" is a name included in the product express column. Currently, NetEase officials have not publicly released a second-generation model under the name "Yuyan 2". The verifiable entity is the Chinese large model "Yuyan" self-developed by NetEase Fuxi and its role-playing version "Character Yuyan". Yuyan first topped the CLUE Chinese language understanding classification list in early 2023 with 11 billion parameters. In June 2026, it was registered as a generative artificial intelligence service by the Cyberspace Administration of China. Its positioning has also shifted from general understanding to plot generation and character dialogue. This report is written based on the current verifiable Yuyan/Miaoqi system, and naming differences will be left to manual review.

• 2. Product background (894 chars)

  Yuyan is a series of Chinese text pre-training large models self-developed by NetEase Fuxi Lab. Founded in 2017, NetEase Fuxi is one of the first domestic research institutions focusing on game and pan-entertainment AI. As of public information, it has published more than 200 AI conference papers and holds more than 500 invention patents. The predecessor of Yuyan is the 5 billion parameter GPT2 generation model self-developed by Fuxi. After a small amount of secondary training, it was transformed into Prefix-LM with strong understanding ability. Subsequent model scale has grown from 100 million parameters to 100 billion MoE architecture, and expanded multi-modal capabilities such as graphics, music, and behavioral sequences. In June 2026, NetEase Fuxi's "Yuyan" large model was registered with the Cyberspace Administration of China. The official clarified that it "does not do general dialogue, but specializes in plot generation and character dialogue", and has spilled its capabilities onto intelligent NPCs in games such as "Backwater".

• 3. Product functions and usage experience (1066 chars)

  Yuyan's core capabilities are divided into two lines. One is universal Chinese understanding. The 11 billion-parameter Encoder-Decoder structure (deep Encoder + shallow Decoder) allows it to exceed human level for the first time in six subtasks of the CLUE classification task, and can be used for text creation, news media, intelligent customer service, etc. The other is role-playing, also known as "CharacterYuyan" - a large character model specially polished for character dialogue. It has won 11 firsts in the 12 subdivisions of CharacterEval, the authoritative benchmark for Chinese role-playing. In the mobile game "Ni Shui Han", the AI ​​NPC driven by jade words can call the player character by name, change the player's attitude according to the player's behavior, and generate "emotional memory" with it; in LOFTER's AI writing activities, a total of 4.13 million stories and 3.3 billion words were generated, with an average of 7 times used per person. Yuyan's large-scale model in the field of medical literature was jointly built with Zhejiang Provincial People's Hospital and fine-tuned with more than 1 million pieces of professional data, covering scientific research scenarios such as document screening, reading, writing and translation.

• 4. Pricing and business model (805 chars)

  The Yuyan system does not directly sell subscriptions to C-sides, but takes the route of "Business Embedded + Enterprise API". "Youlingmiaoqi" for developers has opened the Moa API. The model name is passed to yuyan-plus, and it supports standard parameters such as max_tokens, temperature, top_p, top_k, etc. The upper limit of the conversation history is about 101 rounds. The enterprise API is billed based on the number of calls or customized solutions. Please contact the business for specific quotations. Intelligent NPCs, AI teammates and other capabilities in the game have been integrated into NetEase's own products, and players can experience them for free. This model of "first using AI in the games, music, and education businesses that you are most familiar with, and then spreading your capabilities to others" is completely different from the route that Byte, Baidu, Alibaba, and Tencent have spent heavily on grabbing C-side entrances.

• 5. User reviews (819 chars)

  Player feedback on Yuyan-driven content creation has been positive. LOFTER's AI writing marketing campaign brought in 530,000 users, and the WeChat index reached twice the peak value in the past 30 days, indicating that low-threshold creation has indeed captured the needs of young interest communities. The NPCs in the game that can remember players and change their attitudes have also generally received favorable reviews. The expression recognition of the mobile game "Absolute Deduction" and the interactive word game of "Meet Nishuihan" have remained popular two years after their launch, which is evidence of Yuyan's implementation in real scenes. However, since Yuyan is not a universal assistant open to direct dialogue with the public, ordinary users have almost no independent entrance to use it on the C-side. They experience it more "unconsciously" in NetEase products, which makes it lack the accumulation of public reputation like ChatGPT and Doubao.

• 6. Industry evaluation (834 chars)

  The industry generally regards Yuyan as one of the models that “knows Chinese best”, and its CLUE and FewCLUE rankings are regarded as important milestones in Chinese text pre-training. The "Big Model Home" hot list in May 2026 shows that the industry competition benchmark has shifted from "running scores" to "agent capabilities" and "how long it can work alone." Yuyan is one of the few samples that deeply embeds models into business flows and relies on cost reduction and efficiency rather than burning money to gain access to commercialization. NetEase will invest 17.5 billion yuan in R&D in 2025, with an R&D intensity of 16.6%. AI in game art has replaced more than 50% of repetitive work, and the efficiency of some production links has increased by 300%. This is regarded by analysts as a pragmatic approach to "use AI to save costs" rather than "use AI to grab users". At the same time, NetEase invested in DeepSeek to occupy an early position on the domestic computing power (Ascend) route.

• 7. Controversies and Risks (963 chars)

  The first risk is the ambiguity of naming and positioning. "Yuyan 2" is not an officially announced generational name. The real public form is Yuyan/Miaoqi, and after filing, it is clear that "no general dialogue" is made. This means that it is not in the same track as general bases such as Tongyi, Wenxin, and DeepSeek. Comparing general capabilities will be distorted. The second risk is low openness: Yuyan is not open source or provides a public free dialogue entrance like DeepSeek and Qwen. It mainly serves NetEase's internal and B-side enterprise customers, and the access threshold and uncertainty for external developers are high. The third risk is the scale of the ecosystem. Compared with the large general-purpose model, Yuyan's public discussion and third-party application ecosystem are both smaller, and it is more "game/entertainment vertical" rather than platform-based. The fourth risk is compliance and illusion. Serious scenes such as medical documents have zero tolerance for illusion. Although fine-tuning in the vertical field has alleviated the problem, it is still a core obstacle to production implementation.

• 8. Applicable groups and usage suggestions (721 chars)

  Yuyan/Miaoqi is most suitable for three types of people: teams that need "role playing + plot generation" for games, short plays, film and television animation, education, etc.; organizations that need intelligent assistants in vertical fields (such as medical literature); and enterprise developers who want to take advantage of NetEase's capabilities (using Lingmiaoqi Enterprise API). It is not suitable for individual users who want an open, free, universal conversation assistant, nor for researchers who need open source authority to deploy it themselves. If you want general programming or knowledge question and answer, open models such as Tongyi, DeepSeek, and Zhipu will be more direct choices; if you want a dialogue character with "character, empathy, and remembering the character", Yuyan/Miaoqi is the first echelon in the country.

• 9. Summarize (368 chars)

  Yuyan is a representative work of NetEase's "first use and mature" route of large models, and role-playing and plot generation are its real strengths. The filing in June 2026 gave it a ticket to commercialization; however, the name "Yuyan 2" does not yet have official generational endorsement. It is recommended to confirm during manual review whether it should be changed to "Yuyan/Miaoqi" or marked as a second-generation iteration.

No in-depth report yet

• 

  Richard_Allen_77

  —

  玉言 2 这个命名其实有点模糊，官方公开的是「玉言 / 妙启」，2026 年 6 月备案后明确「不做通用对话」，所以拿它跟豆包、DeepSeek 比通用能力会失真。但如果你的需求是角色扮演、剧情生成、人设不崩的陪伴对话，玉言 / 妙启在 CharacterEval 拿了 11 项第一，确实是国内第一梯队。建议产品页把名字和定位写清楚，免得用户冲着「通用大模型」来结果预期落空。

---

• 

  Brian.Hicks_2023480

  —

  看了 DoNews 那篇「起了大早赶了晚集」的分析挺扎心的。玉言 2023 年就登顶 CLUE 了，参数量却只有 110 亿，和同期 GPT-3 的 1750 亿根本不在一个量级，后来通用赛道也没挤进第一梯队。但换个角度看，网易没硬刚通用基座，而是把玉言沉到游戏、医疗文献这些垂直场景，反而跑通了「先用熟再外溢」的路子。这种清醒未必是退路，更像是一种取舍。

---

• 

  bigtiger267

  —

  2026 年 6 月玉言过了网信办备案，官方明确不做通用对话，等于把定位彻底收拢到剧情生成和角色对话了。

---

• 

  Michelle_Butler5

  —

  逆水寒的门客又给我送奶茶了，离谱。

---

• 

  Alice_Torres520717

  —

  说实话最早玉言登顶 CLUE 的时候我还挺期待的，结果后面两年声量越来越小，直到逆水寒把 AI 门客做出来才又让人眼前一亮。它真正的长板从来不是跟通义文心比参数，而是把角色扮演和剧情生成磨到了国内第一梯队，这种「不抢入口、先在业务里用熟」的打法，反而比盲目烧钱更稳。

---

• 

  Janice_WhiteZ37

  —

  逆水寒的智能 NPC 单局高频互动能到七十轮左右，还带跨局长线记忆，这种陪伴感确实是靠玉言撑起来的。

---

• 

  RWood_88527

  —

  玉言这角色扮演是真能打。

---

• 

  m68o1ws7qd

  —

  我在逆水寒捏了个「江湖友人」，一开始只是图个乐，后来发现它会根据我设定的性格主动找话题，甚至记得上礼拜聊过的事，还会因为我跟别的角色亲近了吃点小醋。这种「情感记忆」真的让人有点细思极恐，但不得不说，玩游戏这么多年第一次觉得 NPC 像个活人。

---

• 

  掠影_8

  —

  在逆水寒里捏了个江湖友人，聊起红楼梦居然主动提宝黛，这记性有点吓人。

---

• 

  刘兰鹏

  —

  丁磊这波「用 AI 省成本」比「用 AI 抢用户」清醒。

---

• 

  greenduck417

  —

  用伏羲的工具做国风游戏场景，输一句「仙剑风格的江南庭院带荷花池」，生成的风格比例都对，只改了两棵树就能用，比找外包省了三千多。

---

• 

  蔡莉

  —

  作为独立游戏开发者，网易伏羲这套工具确实是为国风项目量身做的：全中文、不用翻墙、从场景到 NPC 对话逻辑一条龙。坑也有，欧美科幻风的高模质感不如 MetaHuman，复杂大场景偶尔错乱要自己微调，但对学生和中小团队来说，性价比比找外包高太多了。

---

• 

  Johnny_HartJr

  —

  妙启的角色一致性确实强，聊半小时人设都没崩过。

---

• 

  JosefaCarmona

  —

  玉言医疗文献大模型是和浙江省人民医院共建的，用一百万条专业数据微调，覆盖文献筛选到科研写作，这种垂直打法比硬上通用模型聪明。

---

• 

  Web3DevHill

  —

  学生认证能领半年免费额度，做毕设太香了。

---

• 

  MariaColeman

  —

  捏脸 AI 能读懂「春风拂槛露华浓」，这中文理解确实有两把刷子。

---

• 

  梅花797

  —

  比起通用大模型，玉言走的是游戏这条路，挺务实的。

---

• 

  Ryan_Morris_8823

  —

  玉言 110 亿参数放在今天确实不算大，但人家压根不卷通用对话，专攻剧情和角色对话，在游戏里反而更顺手。

---

• 

  Samuel_WilliamsII

  —

  开源的 yuyan-11b 在 HuggingFace 上能下，想自己跑故事生成的可以试试。

---

• 

  Johnny.AllenK

  —

  企业想接 yuyan-plus 得邮件申请 AK，按量计费，对中小团队门槛不算低，但游戏内玩家是免费体验的。

No reviews yet