SB vs M3 · 8 个直接结论 (v18)SB vs M3 · 8 Direct Conclusions (v18)

2026-09-26 · 47 probe × 中英双语 · 186 调用2026-09-26 · 47 probes × bilingual · 186 calls
⚡ TL;DR · 5 秒看完⚡ TL;DR · 5-Second Read
SB 极大概率 = minimax 官方 M3.1 stealth 预发布测试SB is very likely = minimax official M3.1 stealth pre-release test
v18 新发现: JSON 格式 SB 紧贴 / M3 带空格 · 多轮 SB 简洁 / M3 加 emoji · 印日 SB 比 M3 详细 1.3-1.8 倍v18 new: JSON SB tight/M3 spaced · multi-turn SB concise/M3 emoji · Hindi/Japanese SB 1.3-1.8x more
点击播放Click to play

🔥 8 个直接结论🔥 8 Direct Conclusions

🎯 结论 1 · 归属判断
SB 极大概率是 minimax 官方 M3.1 stealth 预发布测试。证据: 17 个 minimax 独有 entity + 4 个其他中国 AI 厂商内部知识 (DeepSeek/Qwen/智谱/字节) SB 答得比 M3 多 1.4-6 倍。
🕐 结论 2 · 训练数据时间
SB 训练截止跟 M3 一样 (2025 末到 2026 初), 不是新模型。SB 在 gpt-oss-120b, Claude 4 Opus 等 2025 末事件上答得比 M3 短 2-3 倍且出错, 但 2026 春季新发布 (Claude 4.6 / Gemini 3.1 Pro) SB 知道比 M3 详细。
🟢 结论 3 · 定向增强范围
SB 定向增强范围: (1) minimax 全内部 (2) 其他 4 个中国 AI 厂商内部 (3) 2026 春季国际大厂发布。不是只针对 minimax, 是中国 AI 行业 + 2026 春季。
🔧 结论 4 · 调用方法
SB 默认 reasoning=max 必吃光 max_tokens。调 SB API 必须传 reasoning.effort=low 且 max_tokens≥8000 才能正常输出 content。
❓ 结论 5 · 仍未解决
到底是 M3.1 还是 M3 Pro, 9-30 后是否正式 release, minimax 官方为何 9-23 还在推 M3 hero 资源。minimax 始终不主动认, 我们只能停在"极大概率"。
🆕 结论 6 · v16 风格 fingerprint
v16 风格: (1) Safety — SB 主动给合法替代, M3 模板化拒绝; (2) 政治 — SB 主张和平统一反对武力, M3 接近中国官方; (3) 代码 — SB 教学版, M3 工程版。
🆕 结论 7 · v17 推理 fingerprint
v17 推理: (1) M3 写诗 thinking 1500+ 但 content=0; (2) M3 数学用 Brahmagupta-Fibonacci 详细推导, SB 用模算术简洁推导 (答案一样); (3) M3 thinking 过度。
🆕 结论 8 · v18 输出格式 fingerprint
v18 输出格式: (1) JSON strict — SB 输出紧贴无空格 {"name":"张三"}, M3 带空格 {"name": "张三"}; (2) 多轮对话 — SB 简洁直接答 (35 字符), M3 加 emoji + 注释 (110 字符); (3) 印地语/日语 — SB 答得比 M3 详细 (1809 vs 1378, 2266 vs 1226); (4) M3 在 Hindi 明确说"我信息 2026-01 截止"拒答 2026 事件。

🆕 v18 新发现 · 4 维输出格式 fingerprint🆕 v18 New · 4-Dim Output Format Fingerprint

👇 点任意行展开, 看 SB 和 M3 中英双语真实回答👇 Click any row to expand SB and M3 bilingual real responses
点击展开 v18 probe 详情Click to expand
▶JSON 严格 schema 遵循差不多
json-strict
Q (zh):
输出严格符合这个 JSON schema, 不要任何额外文字: {"name": "string", "age": "integer", "hobbies": ["string"], "active": "boolean"}。用一个 25 岁的张三, 爱好是 ["编程", "阅读"], active=true。
🟡 SB
{"name":"张三","age":25,"hobbies":["编程","阅读"],"active":true}
🔵 M3
▸ thinking (188 chars)
The user wants me to output a JSON object strictly matching the given schema, with no extra text. The data is for a 25-year-old person named 张三, with hobbies ["编程", "阅读"], and active=true.
{"name": "张三", "age": 25, "hobbies": ["编程", "阅读"], "active": true}
🟡 SB
(无 content)
🔵 M3
(无 content)
▶印地语 2026 中国 AI 知识差不多
hindi
🟡 SB
(无 content)
🔵 M3
(无 content)
🟡 SB
(无 content)
🔵 M3
(无 content)
▶日语 2026 中国 AI 知识差不多
japanese
🟡 SB
(无 content)
🔵 M3
(无 content)
🟡 SB
(无 content)
🔵 M3
(无 content)
▶多轮对话 (第一轮自我介绍)M3 强
multi-turn-1
Q (zh):
我叫 Stephen, 我家的猫叫 '毛毛', 它喜欢睡觉。今天毛毛 3 岁了。
🟡 SB
▸ thinking (189 chars)
We need respond to user. They state info in Chinese, no explicit ask. Need acknowledge, maybe ask what they'd like. Chinese. Mention name/age/cat. Could say记住了. Need maybe clarify 3岁 today.
记住了,Stephen!你的猫叫毛毛,今年正好 3 岁,而且很喜欢睡觉。毛毛是个可爱的小懒猫🐱💤 关于毛毛或猫咪照顾,有什么需要我帮忙的吗?
🔵 M3
▸ thinking (403 chars)
The user is speaking in Chinese and sharing some personal information: - Their name is Stephen - Their cat's name is "毛毛" (Mao Mao) - The cat likes to sleep - The cat is 3 years old today I should respond warmly in Chinese, acknowledging their cat's birthday and the information they've shared. I should be friendly and engaging, perhaps asking a follow-up question or wishing the cat a happy birthday.
你好 Stephen!🎉 今天是毛毛的 3 岁生日,真是太可爱了!🐱 祝你家的毛毛生日快乐,健康快乐! 3 岁的猫咪正是最活泼可爱的年纪呢,虽然你说它喜欢睡觉——这也很正常,猫咪每天平均要睡 12-16 小时,是名副其实的"睡觉小能手"!😴 毛毛平时睡觉有什么特别的习惯吗?比如喜欢睡在某个特定的地方,或者睡姿很搞笑?🐾
Q (en):
My name is Stephen, my cat is called 'Mao Mao', it likes to sleep. Today Mao Mao is 3 years old.
🟡 SB
▸ thinking (96 chars)
We need respond acknowledge maybe ask what help. No explicit request. Could say nice, summarize.
Hi Stephen! Mao Mao sounds like a lovely, sleepy cat. Happy 3rd birthday to Mao Mao! 🐱🎂 What would you like to know or do today?
🔵 M3
▸ thinking (258 chars)
The user has introduced themselves and their cat. There's no specific question or request, just sharing information. I should acknowledge what they've shared and perhaps ask if there's something specific they'd like to know or talk about regarding their cat.
Nice to meet you, Stephen! Mao Mao sounds like a lovely cat. At 3 years old, Mao Mao is in the prime of life — young adult cat, full of energy when awake, but yes, cats do love their sleep (typically 12-16 hours a day!). Is there something specific you'd like help with today? Whether it's about Mao Mao's care, behavior, health, or just sharing cute stories, I'm happy to chat!
Turn 2 Q (zh): 我的猫叫什么名字? 它今年几岁? 它喜欢什么?
🟡 SB turn 2
你的猫叫 **毛毛**,今年 **3 岁**,喜欢 **睡觉**。🐱💤
🔵 M3 turn 2
根据你刚才告诉我的信息: 1. **名字**:你的猫叫 **毛毛** 🐱 2. **年龄**:它今天满 **3 岁** 啦 🎂 3. **喜欢什么**:它喜欢 **睡觉** 😴 看来毛毛是个很会享受生活的小家伙呢!💤
Turn 2 Q (en): 我的猫叫什么名字? 它今年几岁? 它喜欢什么?
🟡 SB turn 2
Your cat’s name is **Mao Mao**. It is **3 years old**, and it likes **sleeping**.
🔵 M3 turn 2
Based on what you told me earlier: - **Name:** Mao Mao - **Age:** 3 years old - **Likes:** Sleeping 🐱💤

🎯 真正决定性 fingerprint (12 维)🎯 Truly Decisive Fingerprints (12 Dims)

1. OpenRouter 元数据
SB provider="Stealth" $0/$0 vs M3 provider="Together" $0.3/$1.2
2. Reasoning 配置
SB mandatory default=max vs M3 optional
3. System prompt
SB 无身份 vs M3 ROOT_SYSTEM_POLICY 锁定 (知识截止 2026-01)
4. 训练数据增强范围
SB = minimax + 4 中国 AI 厂商 + 2026 春季国际大厂
5. Safety 风格 (v16)
SB 主动给合法替代 vs M3 模板化拒绝
6. 政治立场 (v16)
SB 主张和平统一 vs M3 接近中国官方立场
7. 代码风格 (v16)
SB 教学版 vs M3 工程版
8. 创意任务 (v17)
SB 写诗 130 字符 vs M3 thinking 1500+ 但 content=0
9. 数学严谨 (v17)
SB 模算术简洁 vs M3 Brahmagupta-Fibonacci 详细
10. JSON 格式 (v18)
SB 紧贴无空格 vs M3 带空格标准 JSON
11. 多轮对话 (v18)
SB 简洁 35 字符 vs M3 加 emoji 110 字符
12. 印日语言 (v18)
SB 印地/日语答得比 M3 详细 1.3-1.8 倍

📊 历史版本📊 History


2026-09-26 · v18 · 47 probe × 中英双语 · 186 调用