事情是这样的。
装上 Kilo Code,搜 free,弹出来 9 个模型。

Auto Free、Step 3.7 Flash、Laguna M.1、Nemotron 3 Ultra……
懵了。
选哪个?哪个能用?哪个有坑?
翻了圈中文社区,没人系统讲过。那我来。
一句话结论
日常主力用 StepFun Step 3.7 Flash,复杂任务换 Laguna M.1,敏感代码一个都别用。
原因后面展开。
先看这 9 个模型长什么样:

Kilo Code 把免费模型分了两层:
-
推荐区(上面那 3 个):官方筛过,相对靠谱 -
Kilo Gateway 区(下面那 6 个):合作厂商直供,良莠不齐
先讲推荐区,再说 Gateway 区。最后说哪些打死不能碰。
第一步:认准 Step 3.7 Flash,大多数时候用它就够了
这是截图里当前选中的那个,也是信息最全的一个。
|
|
|
|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
翻译成人话:196B 的大模型,但每次只跑 11B,又强又快。256K 上下文,能一口气读完一个中型项目的全部代码。能看图,能看视频。
在 AI 编程工具里,免费模型带多模态的,一只手数得过来。
怎么用最划算:
简单补全 → 推理开 Low,秒出结果 复杂重构 → 推理开 High,让它多想一会 有设计稿 → 直接贴图,让它写前端代码

第二步:Laguna M.1 留着干重活
Poolside 出的。Poolside 这家公司只做一件事:AI 写代码。
在免费区里,Laguna 有两个:M.1 和 XS.2。M.1 是大的,XS.2 是小号的。
规律很简单:代码逻辑复杂的用 M.1,只想快速补两行用 XS.2。
但 XS.2 在 Gateway 区,说明官方对它的信心不如 M.1。
第三步:Auto Free 最方便也最危险
Auto Free 排在推荐区第一个,Kilo 的自动路由——你不用选,它帮你挑当前最优的免费模型。
方便是真方便。但截图里有一个细节,很多人扫一眼就过了:
⚠️ 数据可能会用于训练
这行字在 Step 3.7 Flash 的详情面板底部,但它不仅适用于 Step 3.7 Flash。
Auto Free 的请求可能被路由到任何厂商的服务器。你的代码去到了哪里、谁在看、被用来干什么——你不知道。
哪些情况可以用 Auto Free:
-
写 Demo、学习项目、开源代码 -
不含任何公司信息的个人练习
哪些情况绝对不要碰:
-
公司业务代码 -
包含客户数据的项目 -
加密逻辑、认证模块、支付相关 -
配置文件中出现过 password / token / key 的地方
一句话记牢:敏感代码用免费模型 = 代码裸奔。

第四步:Gateway 区的 6 个,按需取用
然后才是 Gateway 区。
Cohere North Mini Code
名字里 Mini 和 Code 都在,轻量、代码专精。Cohere 企业级 NLP 的底子,不会太差。日常辅助够用。
Free Models Router
不是具体模型,更像一个负载均衡入口。和 Auto Free 的区别:Auto Free 是智能选优,Router 是随机分发或轮询。稳定性有余,质量没保证。
NVIDIA 三兄弟:Nano Omni / Super / Ultra
截图里 NVIDIA 一口气占了三席。像吉祥三宝。
|
|
|
|---|---|
|
|
|
|
|
|
|
|
|
有意思的是:NVIDIA 这三个全免费。明摆着推广自家生态,先让你免费用,习惯了你自然会看他们的付费模型。
⚠️ NVIDIA 免费端点的数据使用条款和 Step 3.7 Flash 一样——数据可能被用于训练。 甚至更严格。根据 NVIDIA API 试用条款,仅限试用、不可用于生产环境或处理敏感数据。
除非你对 NVIDIA 生态有特殊需求,否则这三个不是最优选。
Poolside Laguna XS.2
Laguna 的小号版。速度比 M.1 快,能力比 M.1 弱。写几行简单代码还行,别让它干架构设计的活。
给不同人的配置方案
学生 / 自学者
主力:Step 3.7 Flash。图多就开多模态,代码多就开推理 High。完全够用。
个人开发者
Step 3.7 Flash(日常)+ Laguna M.1(复杂任务)。两个轮换,速率限制来了就切。
有保密要求的在职开发者
⚠️ 别省这点钱。 公司代码丢不起。老老实实用付费模型,免费模型只用来查文档、学新框架。
技术博主 / 开源贡献者
随意用。公开代码本来就不怕看,免费的不用白不用。顺手还能反馈一下模型效果,推动生态变好。
三个打死不能犯的错
错误一:敏感项目用 Auto Free
你的代码去了哪里、经过哪家厂商的服务器、会不会被存下来——三个问号。Auto Free 就是开盲盒,但你没拿到礼物,是送出去了礼物。
错误二:生产环境依赖免费模型
免费模型是上游厂商免费提供的。今天有,明天可能就没了。速率限制随时会变。你线上的工作流不能建立在"别人免费给你用"的基础上。
错误三:把 API Key 或密码贴进对话
截图里白纸黑字写着"数据可能会用于训练"。你贴进对话的 key、token、密码,理论上都可能被存下来。这不是猜的,是条款里的。
总结
Kilo Code 的免费模型确实能打。尤其是 Step 3.7 Flash——多模态、256K 上下文、MoE 高效架构,这三个标签放在一起,在免费圈几乎没有对手。
但"免费"的前提是数据可能被用于训练。这一点,截图里写了,我也再说一遍。
用之前,先想清楚:你写的这行代码,能接受被别人看到、被存下来、被拿去训练模型吗?
能,就放心用。不能,就老老实实花钱。
就这么简单。

Here's the situation.
Install Kilo Code, search for "free," and 9 models pop up.
Auto Free, Step 3.7 Flash, Laguna M.1, Nemotron 3 Ultra...
Completely stumped.
Which one to pick? Which ones actually work? Which ones have pitfalls?
Scrolled through the Chinese community, no one has systematically covered this. So I will.
One-sentence conclusion
For daily main use, go with StepFun Step 3.7 Flash; switch to Laguna M.1 for complex tasks; don't use any of them for sensitive code.
Reasons will be explained below.
Let's start by taking a look at what these 9 models look like:
Kilo Code splits the free models into two tiers:
Recommended section (the top 3): Officially filtered, relatively reliable
Kilo Gateway section (the bottom 6): Directly supplied by partner vendors, mixed quality
First the Recommended section, then the Gateway section. Finally, which ones you should absolutely never touch.
Step 1: Stick with Step 3.7 Flash; it's enough most of the time.
This is the one currently selected in the screenshot, and also the one with the most complete information.
Project
Data
Developer
StepFun
Total parameters
196B
Activation per request
~11B (MoE sparse activation)
Context window
262K tokens
Supported modalities
Text + Image + Video
Inference mode
High / Medium / Low (adjustable)
Designed for
Coding, Agent workflows, structured output, long-text handling
In plain English: A 196B parameter model, but only 11B activated per request, making it both powerful and fast. 256K context window means it can digest an entire mid-sized project's codebase in one go. Handles images and videos too.
In AI coding tools, free models that support multimodality are rare enough to count on one hand.
How to use it most cost-effectively:
Simple completion → Set inference to Low for instant results. Complex refactoring → Set inference to High to let it think longer. With design mockups → Paste the image directly and have it write frontend code.
Step 2: Keep Laguna M.1 for heavy lifting
Made by Poolside. This company does only one thing: AI code generation.
In the free tier, there are two Lagunas: M.1 and XS.2. M.1 is the larger one, XS.2 is the smaller.
The rule is simple: Use M.1 for complex code logic, XS.2 when you just need to quickly add a couple of lines.
But XS.2 is in the Gateway section, indicating the official team has less confidence in it compared to M.1.
Step 3: Auto Free is the most convenient and also the most dangerous
Auto Free sits first in the Recommended section. It's Kilo's automatic router—you don't have to choose; it picks the current best free model for you.
It's genuinely convenient. But there's one detail in the screenshot that many people gloss over:
⚠️ Data may be used for training
This line appears at the bottom of the Step 3.7 Flash detail panel, but it applies to more than just Step 3.7 Flash.
Auto Free requests may be routed to any vendor's servers. Where your code goes, who sees it, what it's used for—you have no idea.
When you CAN use Auto Free:
Writing demos, learning projects, open-source code
Personal practice without any company information
When you should ABSOLUTELY avoid it:
Company business code
Projects containing client data
Encryption logic, authentication modules, payment-related code
Anywhere password / token / key appears in config files
Remember this in one sentence: Using free models for sensitive code = leaving your code completely exposed.
Step 4: The 6 in the Gateway section, use as needed
Then comes the Gateway section.
Cohere North Mini Code
"Mini" and "Code" are both in the name, lightweight and code-specialized. Backed by Cohere's enterprise NLP foundation, it shouldn't be bad. Good enough for daily assistance.
Free Models Router
Not a specific model, more like a load-balancing gateway. Difference from Auto Free: Auto Free intelligently selects the best, while Router randomly distributes or round-robins. Stability is fine, but quality isn't guaranteed.
The NVIDIA trio: Nano Omni / Super / Ultra
NVIDIA occupies three spots at once in the screenshot. Like the lucky trio.
Model
One-liner
Nano Omni
Smallest, fastest speed, weakest capability
Super
Middle tier, neither here nor there
Ultra
Largest, but code capability doesn't rank in the top 90
Interesting fact: All three NVIDIA models are completely free. Clearly pushing their own ecosystem—let you use them free first, and once you're hooked, you'll naturally check out their paid models.
⚠️ NVIDIA's free endpoint data usage terms are the same as Step 3.7 Flash—data may be used for training. Even stricter. According to NVIDIA API trial terms, it's for trial use only and cannot be used in production or for processing sensitive data.
Unless you have specific needs for the NVIDIA ecosystem, these three aren't the top choice.
Poolside Laguna XS.2
Smaller version of Laguna. Faster than M.1 but weaker. Fine for writing a few lines of simple code, but don't let it handle architecture design.
Configuration plans for different users
Students / Self-learners
Main: Step 3.7 Flash. Enable multimodal for image-heavy tasks, set inference to High for code-heavy tasks. More than enough.
Independent developers
Step 3.7 Flash (daily) + Laguna M.1 (complex tasks). Rotate between the two; switch when rate limits hit.
Working developers with confidentiality requirements
⚠️ Don't skimp here. You can't afford to lose company code. Just use paid models honestly; reserve free models only for looking up docs and learning new frameworks.
Tech bloggers / Open-source contributors
Use freely. Public code isn't afraid of being seen anyway; free models are a waste if not used. Plus, you can provide feedback on model performance and help improve the ecosystem.
Three mistakes you should absolutely never make
Mistake 1: Using Auto Free for sensitive projects
Where your code goes, which vendor's servers it passes through, whether it gets stored—three question marks. Auto Free is like opening a blind box, except instead of getting a gift, you're giving one away.
Mistake 2: Relying on free models for production environments
Free models are provided at no cost by upstream vendors. They might be here today and gone tomorrow. Rate limits can change anytime. Your live workflows can't be built on the premise of "others giving you free access."
Mistake 3: Pasting API keys or passwords into the chat
The screenshot explicitly states in black and white: "Data may be used for training." The keys, tokens, and passwords you paste into the chat could theoretically be stored. This isn't speculation; it's in the terms.
Summary
Kilo Code's free models are genuinely capable. Especially Step 3.7 Flash—with multimodal support, 256K context, and MoE efficient architecture, those three tags together make it nearly unmatched in the free tier.
But the catch with "free" is that data may be used for training. It's stated in the screenshot, and I'm saying it again.
Before using it, think clearly: Can you accept your code being seen by others, stored, and used to train models?
If yes, use it with confidence. If no, just pay honestly.
That's all there is to it.