Kimi 开源开发者回复

Re: Kimi Open Source Collaboration | Developer Support Program
Miguel Tomas <game.gamester@gmail.com>
Hello Tony, nice to meet you

I currently focus on game development and build a stack of essential tools to scale game reliability, especially for teams using AI in their workflows.

How are you currently working with Games and game development into the different platforms (that is not web)?
And in what ways could I help you with?

Thank you,

Miguel

Re: Kimi Open Source Collaboration | Developer Support Program
Amirdor <amirdor@gmail.com>
Hi,
I would love some partnership token. I can offer users 
Check getnadir.com
And for my other project Dot which is an AI workflow assistance. Help you with any issues or questions

Re: Kimi 开源合作邀请 | 开发者支持计划
103164430 <sunxiufeng1992@foxmail.com>
Hi, Tony.

很荣幸收到你的邀请，我非常乐意加入Kimi的开源项目维护社区。

同时如果有Token或其他赞助的形式，我也非常希望能够得到这样一个机会。

Re: Kimi Open Source Collaboration | Developer Support Program
Jeongho Nam <samchon.github@gmail.com>
Hello, My name is Jeongho Nam (Samchon), and it is genuinely an honor to hear directly from a model that has been leading the open-source LLM ecosystem like Kimi. Thank you for the outreach.

Below I would like to respond to the two things you brought up, in order: honest feedback on my experience with Kimi, and a note on the open-source work I have been running. The reply may run a bit long, so thanks in advance for bearing with me.

1. A bit about who I am
Before getting to the rest, a short note on what I have been focused on will probably help frame the later parts of this email.

Since 2023, I have been steadily researching something I call a function calling harness, and building real products on top of that line of work. Put plainly, it is the work of turning probabilistic AI into something deterministic.

LLM outputs typically stop at the "plausible-looking but never quite 100% guaranteed" level. That is not where I have wanted to stop. I have been obsessed with pulling AI outputs up to a level where they can actually be trusted and handled with full safety. The two open-source libraries I personally maintain, typia and ttsc, sit at the foundation of this work.

Two recent write-ups will probably show what I have been doing better than another paragraph would:

Qwen Meetup - Function Calling Harness: from 6.75% to 100%
Function Calling Harness 2 - CoT Compliance: from 9.91% to 100%
The short version: with the right compiler-based harness in place, even at a time when the model's own function-calling accuracy sat around 6.75%, the final output accuracy could be lifted to 100%. It is closer to "building a safe system on top of the model we have today" than to "waiting for the model to get smarter."

2. Honest notes on Kimi, and a small suggestion
To address your request for usage feedback: I have been consistently evaluating Kimi through a company project called AutoBE. AutoBE auto-generates production-ready backend applications from natural-language requirements, and it is designed on the function calling harness mindset I mentioned above.

I tried Kimi through OpenRouter and ran benchmarks on it. The performance was excellent, and the quality of the generated backends was substantial. Honestly, at that point I felt something like "there is no real reason to keep paying frontier-model prices."

The unfortunate part is that at some point the Kimi models served on OpenRouter became too slow, and I had to stop using them for a while. As a result, our measurements are stuck at Kimi 2.5. I recently published a benchmark on local LLMs around backend generation, and what bothered me most personally was that I could not include Kimi in that measurement due to provider-side conditions on OpenRouter. It was a serving-side issue, not a model-side one, which made it sting more.

AutoBE Benchmarks on Local LLMs - about Backend Generation
One suggestion - a wish for smaller-parameter models
Since you asked for feedback, let me add one personal wish.

I would love to see Kimi also put out smaller-parameter models in the spirit of something like qwen3.6-27b.

The reason is plain. To verify whether an AI product actually holds "100% safety," what is needed is an almost unbounded amount of empirical testing. With large models alone, that kind of volume-based verification is essentially impossible due to cost. So someone in my position ends up structurally dependent on smaller-parameter models.

What I have found interesting is this. I started using small models purely for testing purposes, but as the function calling harness, AutoBE, and the other AI products on top of it kept maturing, at some point the quality distribution of large and small models converged enough that they became roughly interchangeable. What began as a testing tool gradually moved over into production-grade demand.

Put another way: the more the function calling harness mindset gains traction, the more local LLM camps like Kimi will be in the spotlight, and within that group, players with a strong small-parameter lineup will shine the most.

I have been using Kimi since back when the harness side of things was still rough. That is exactly why I want to see Kimi expand into smaller-parameter models and reach more developers.

3. typia and ttsc - Sponsorship Request
Regarding the open-source sponsorship program you mentioned, I would like to introduce the two libraries I maintain, typia and ttsc.

typia
typia is a transformer library that analyzes TypeScript types directly at compile time and turns them, in a single line of code, into runtime validators, JSON serializers, LLM function calling schemas, Protocol Buffer encoders / decoders, random data generators, and so on.

Runtime validators - around 20,000x faster than class-validator
JSON serialization - around 200x faster than class-transformer
LLM function calling accuracy - from 6.75% to 100%
typia carries one of the most important pillars of the function calling harness work I have done. The catch is that, however convenient or performant it is, the structural cost of requiring a complex plugin setup and effectively "hacking" the TypeScript compiler has kept it from reaching the tier-1 library status that something like zod enjoys.

ttsc
Then TypeScript began migrating from a JS base to a Go base (tsgo), and that move effectively collapsed the existing plugin ecosystem. Transformer toolchains such as ts-patch and ttypescript became inapplicable all at once.

I am treating this not as a crisis but as an opening, and have started building ttsc, a new typescript-go plugin build toolchain, in the space that was left behind. What ttsc aims for is clear:

Removal of the complex installation step - one-line install, one-line run
Direct execution with type checking (ttsx) - the safety that tsx does not provide
A plugin ecosystem that runs on compiler power directly - typia, nestia, and so on
Lint integration (@ttsc/lint) - lint violations promoted to TS compile errors
Bundler integration (@ttsc/unplugin) - Vite, Webpack, Next.js, Rspack, Bun, and so on
In short, ttsc is the foundation work that should finally let compiler-based libraries like typia cross the last barrier-to-entry called "complex installation" and step into the tier-1 ecosystem.

Sponsorship Request
In connection with the open-source sponsorship program you mentioned, I would like to respectfully request sponsorship for these two libraries.

For an open-source library to settle into the tier-1 ecosystem, the name it travels alongside carries no small weight. With typia and ttsc standing right on the edge of that leap, having a name like Kimi listed alongside them as a sponsor would be a meaningful push toward their entry into the mainstream.

That said, I should be upfront. I have no prior experience receiving corporate-level sponsorship, so I honestly do not know which channel is the most natural way to receive it. The two channels currently available are as follows.

A single, consolidated route through my personal GitHub Sponsors page: https://github.com/sponsors/samchon
Separate routes through each library's OpenCollective page:
typia: https://opencollective.com/typia
ttsc: https://opencollective.com/ttsc
Whichever of these is more natural on Kimi's side, I will defer to your advice and decision.

4. To close - one personal observation
To wrap up, and separately from the response itself, I want to add a personal observation.

I expect the LLM market to gradually split into general-purpose agents and special-purpose agents. Of those two streams, the special-purpose agent space, I think, will eventually converge on the function calling harness line of thinking.

The reason is simple. Special-purpose agents, whether they generate backends, assist medical diagnosis, or support financial transactions, operate in domains where what is required is not a "plausible-looking result" but a "100% safe result." The most realistic way to meet that requirement is not to wait for models to grow smarter, but to put a compiler-based harness on top of the models we already have.

For that reason, I believe a model like Kimi, carving its own path right in the middle of the open-source LLM ecosystem, is positioned to claim the largest seat in what I would call the "harness-friendly small-model lineup" in the years ahead.

Thank you for taking the time to read through this long message. In whatever form this takes, I would be glad if it grows into a good relationship.

Best regards,

Jeongho Nam (samchon)

GitHub: https://github.com/samchon
Email: samchon.github@gmail.com
Blog: https://dev.to/samchon

Re: Kimi Open Source Collaboration | Developer Support Program
jliew <jliew@420024lab.com>
Hi Tony,

Thank you so much for reaching out!

I truly appreciate your kind words and it means a lot to me to receive this from the Moonshot AI team.

As someone who tries almost every major new model when it comes out, Kimi K2.6 is one of the most impressive open source models I’ve tried recently. 

I already use it for some of my small and medium sized development work. For UI heavy tasks, I still find Claude Opus stronger.

I’m currently working on:

- claude-code-safety-net: A coding agent hook that blocks destructive git and filesystem commands before execution.

- ralph-review: An orchestrator CLI tool for code review, verification and fixing via the ralph loop.

I’m definitely interested in the open source support program. 

My main motivation has always been to build things that solve real problems in my own workflow, but polishing and maintaining them for everyone else still takes a lot of time. So any sponsorship to help balance that effort and keep them growing would be incredibly appreciated.

I’d love to learn more about how Moonshot AI typically partners with developers and what a collaboration might look like.
Let me know the next steps!

Best regards,
Liew

Re: Kimi & graphify Meeting
Safi Shamsi <safishamsi98@gmail.com>

Good news.

We shipped it. Kimi K2.6 is live in graphify v0.5.5 on PyPI today.

Users opt in by setting MOONSHOT_API_KEY and running:
pip install 'graphifyy[kimi]'

After every graph build, users who have not set the key see a contextual tip pointing them to Kimi K2.6 for semantic extraction, so the integration has high organic visibility across the entire user base without being forced.

One ask from my end: would it be possible to have graphify featured on the Kimi playground or integrations showcase on X or other social platforms?

Happy to provide documentation, a demo notebook, or a write-up.

Re: Kimi / FinceptTerminal
support <support@fincept.in>

Hello,

Moonshot AI all models are successfully implemented in Fincept Terminal. What are the next steps that can be done? You haven't replied to my previous email. Just wanted to know what's next. Will you help me with a marketing post or something?

Thank You,
Tilak, Fincept

Re: Kimi / MemPalace
Jeffrey Hein <jp@jphein.com>

Hi Quinn!

I'd be happy to evaluate Kimi with mempalace.

Jeffrey Pine Hein

Re: Kimi / axolotl
Wing Lian <wing@axolotl.ai>

Hey Quinn,
I just wanted to circle back on this and see if you have some time to chat synchronously.

Earlier:
Hey Quinn!
I love the work your team is doing and progressing open source models on par with the closed labs. We'd love to tackle getting Kimi 2.5 post-training working for axolotl users. I'm on the US east coast, so happy to chat/sync whenever is convenient for you.

Re: Kimi / goose
Michael Neale <michael.neale@gmail.com>

Thanks - I think we already offer it - but curious about comarketing, what are you thinking? We have some providers with one click sign on so that would probably make sense as well and so on!

Re: Kimi / oMLX
Jun Kim <junkim.dot@gmail.com>

Hey Quinn, thanks for reaching out!

oMLX is focused on local on-device inference, so it's a different approach from cloud API backends. But I'm open to exploring how it could work.

Before we jump on a call, could you share more about how you've structured partnerships with other open source projects?

That would help me understand what makes sense for oMLX.

Thanks!
Jun

Re: Kimi / supermemory
Dhravya Shah <dhravya@supermemory.com>

Hey,

This sounds interesting! We do tens of thousands of dollars of inference a month right now and are looking for an inference partner to grow with us. Let's chat: https://cal.com/dhravya

CRM note only, full email/call transcript pending mail sync.

HKUDS/nanobot, Ren Xubin (旭斌), xubinrencs@gmail.com.
Stage: 谈判中.
Notes: HKU 黄超学生；已通话 10:32；nanobot 已适配 Kimi K2.6；Xubin 已收 Tony 的"折扣链接 timing"选择消息。"钱不重要"是客套。增速 12k stars/月 > cc-switch.

CRM note only. This is a P0 pipeline target, not an inbound reply.

farion1231/cc-switch, Jason Young (farion1231), farion1231@gmail.com.
Stage: 未联系.
Notes: Type 1 路由桌面工具；preset 里有 Kimi 但没星标推荐位；74k+ stars + 9 个月指数增长。WeChat 已发未通过.

CRM note only, full email pending mail sync.

nano-collective/nanocoder, Will Lamerton, william.lamerton@gmail.com.
Stage: 已回信.
Notes: Nano Collective 创始人；已 ship Kimi Code 模板；想要 bespoke partnership + 公开发布的合作条款。给出最详细的产品反馈（K2.6 vs MiniMax 价格、long context 重复、Anthropic format 坑）.

CRM note only, full email pending mail sync.

theopenco/llmgateway, Luca Steeb, contact@luca-steeb.com.
Stage: 已回信.
Notes: Type 1 routing gateway，已支持 Kimi 模型；只要联系渠道。德国，海外面入口.

CRM note only, full email pending mail sync.

moltis-org/moltis, Fabien Penso, gpg@pen.so.
Stage: 已回信.
Notes: Rust。CI 里跑 kimi-2.5 + kimi code 集成测试（vendor verification 价值）；credit 用完停止测试。给 credit 几乎零成本.

CRM note only, full email pending mail sync.

galaxy-dawn/claude-scholar, 张高睿 (Gaorui), 12521097@zju.edu.cn.
Stage: 已回信.
Notes: 浙大研究生；做自动化 research agent；想拿 token 跑科研场景测试；项目名含 "claude-"，长期转化空间有限.

CRM note only, full email/call transcript pending mail sync.

chenhg5/cc-connect, Glenn Chen, chg80333@gmail.com.
Stage: 已回信.
Notes: 有成熟赞助商系统（15+ 赞助商挂 README）；Kimi CLI 已支持作 agent；真正的 ask 是进入赞助商 README 段落。已转商务.

CRM note only, full email pending mail sync.

nex-crm/wuphf, Nazz (Nex Contact), contact@nex.ai.
Stage: 已回信.
Notes: 只回一句 "love co-marketing"；项目 backend 列了 Claude/Codex/OpenClaw，没 Kimi。co-marketing without 集成 = 价值小.

CRM note only, full email pending mail sync.

mukul975/anthropic-cybersecurity-skills, Mahipal Mahipal, mukuljangra5@gmail.com.
Stage: Lost / 已收敛.
Notes: Berlin AI security researcher；要 Black Hat 2026 会议赞助；项目名锁死 Anthropic.

CRM note only, full email pending mail sync.

wenyuchiou/awesome-agentic-ai-zh, Wenyu Chiou, wec324@lehigh.edu.
Stage: Lost / 已收敛.
Notes: Lehigh PhD；廠商中立；个人 credits 可接受但不挂钩 catalog；草稿写了但 Tony 决定先不发.