---
title: "话说出了好多新模型啊。。。 - World Discussions - Yumina Community"
description: "有没有哪位大佬详细测评一下模型的实际生成质量呢，如Gemini 3.7和3.8在文字逻辑生成上是否比其他Gemini模型遥遥领先呢?一搜新闻就是遥遥领先，我玩的时候也试一试看内容呢，看看在有大量指令填充内容下，其内容量和逻辑性和常用的Gemini 3的区别大不大。。毕竟相对于Grok系列的高消耗，Gemini性价比还是不错的，DeepSeek个人总结就是狗都不用，现在消耗高了更不用说了。"
url: https://yumina.io/app/community/thread/8184d07f-f45f-4227-a31b-e29832a4d270
lang: zh
---

[Community](https://yumina.io/app/community) · [World Discussions](https://yumina.io/app/community/world-discussions)

# 话说出了好多新模型啊。。。

[大大爆肝](https://yumina.io/@zxyw) · 2026-09-08

有没有哪位大佬详细测评一下模型的实际生成质量呢，如Gemini 3.7和3.8在文字逻辑生成上是否比其他Gemini模型遥遥领先呢?一搜新闻就是遥遥领先，我玩的时候也试一试看内容呢，看看在有大量指令填充内容下，其内容量和逻辑性和常用的Gemini 3的区别大不大。。毕竟相对于Grok系列的高消耗，Gemini性价比还是不错的，DeepSeek个人总结就是狗都不用，现在消耗高了更不用说了。

## 3 replies

1. [🐱](https://yumina.io/@stay0325) · 2026-09-09 目前测了这些，其他的晚点/明天吧 模型测试 （全部都是使用同一张角色卡测试的结果）（8k+token角色卡、含有变量、面板、1v1角色对话） 1. Mistral Nemo -第一轮就直接丢失状态面板 -narrator输出内容的末尾会有乱码（第一轮就出现） -回复越来越短 -已读乱回 -角色OOC -一个回复0.2~0.4蘑菇币 2. Gemini 2.5 Flash Lite -状态面板时而消失时而出现 -会出现英文 -过度文学化、冗余严重、推进效率低 -一个回复1~2.4蘑菇币 3. Mitral Small 3.2 -会出现英文版的逗号在中文回复里 -模型有点容易busy所以要等/一直重试 -一个回复0.9~1.2蘑菇币 4. GLM 5.3 Flash -会出现英文 -会少放标点符号 -一个回复1.3~1.9蘑菇币 5. Gemini 2.5 Flash -会出现少量英文单词 -总体来说很不错 6. Qwen3 VL 235B -人设设定女但narrator说“我”是男的 -生成内容较慢 -轮数多了内容会开始重复 -末尾会出现（场景进行中） -一个回复1.7~2.2蘑菇币 7. Claude 3 Haiku -回复较短 -全文英文逗号没有space -会重复user讲的话 -narrator的回复有概率出现整个内容都是user在讲什么、直接替代user来说话、做事、推剧情 -一个回复3.4~3.7蘑菇币 8. Hunyuan 3 -生成内容较慢 -状态面板内容不全 -逗号是这样的، -轮数5就乱码了 9. Gemini 3.7 Flash -整体使用感不错，值得一试 -一个回复6~9蘑菇币
2. [Swuell](https://yumina.io/@swuell) · 2026-09-12 You’re honestly better off using a Chinese llm instead of trying to use an English one. Since the majority of the context the llm is fed is English and English related syntax. The upgrades you’re talking about is generally about this and more specifically about other lllm uses that’s not really related to using it as a chat bot. P.S. the higher version you go the more nsfw is blocked in terms of Gemini. And the more soft lock and triggers you encounter. Deepseek was good—honestly—the issue is the context and usages that were sent to the model which the model also trained and fed on… so now deepseek is where we are at. A lot of English players also used deepseek before so that actually could have—and probably did— mess with the Chinese syntax and usage. So now it’s a mix of Chinese and English. The only way around that would be to either separate the model or have it compartmentalize each individual language to its own segment, but I don’t think they’re too worried about doing that when the main use isn’t even fit chat bots. So in terms of what they want to do with it deepseek works fine as it is.
3. [Swuell](https://yumina.io/@swuell) · 2026-09-12 Have you tried doing the same thing but with English and then compare the two? Honestly that’d be the best way to see what the differences are and why. Edit: in terms of Qwen the usage is technically right since he/him is still considered neutral. And female is often not used as much in online novels. So based on what the llm model training data is, you’d get more male related pronouns. Though in general persona wise that is annoying.
