#Moonshot (Kimi)
floral stratus
noble hare
THis is what my agent recommended when i ran out of claude subscription and started using kimi2.5
Updated Agent Recommendations
Agent -- Model -- Why
planner -- meta/llama-3.3-70b-instruct -- Best of what's available
coder -- qwen/qwen2.5-coder-7b-instruct -- Verified coding specialist
main -- moonshotai/kimi-k2.5 -- Keep current (200K ctx)
research -- meta/llama-3.1-405b-instruct -- 405B for heavy research
What do you think ??
lavish berry
7b model as coder seems very bad
noble hare
i realise kimi2.5 for everyhting is decent enough
alpine crane
⚠️ API rate limit reached. Please try again later.
Does anyone know what's up with that? I got it literally on the first request after 6 hours of not sending a single one. And that's not the first time either. They all auto-retry successfully, but it is irritating.
noble hare
It is a shared model, so if the traffic is too high your query queues and wait time is 60seconds. Also the processing is slow, they probably also started puting a rate limiter to 40rpm or something that limits how much power you can pull.
Cant run kimi on nividia nim free tier anymore
**sorry i assumed nvidia case, kimi2.5 from moonshot worked worked for me
high plank
I've spent $27 with 1k requests and 84.6M Tokens via OpenRouter and Kimi K2.5.
alpine crane
ok, a couple of days later, and here's my verdict: I can't really recommend Moonshot. The model is great, but they dock your quota by 1%p for every API request that gets rejected---even if your quota is more than fine and it's the server being overloaded. That's not ok---I paid for that quota. And while I can understand that there may be times when the server is overloaded and I have to wait a moment or two, getting punished on top of being refused service is too much.
Fair would be to credit quota when the server is overloaded and the paid service cannot be provided.
unborn path
Hi, I tried to use Kimi k2.5 in my browser automation agent but it stopped working suddenly during the execution. How can I fix it? Even Minimax doesn't stop at that time
humble nebula
anyone else having issues using 2.6 and tools not working?
balmy pivot
I just purchased a subscription for Kimi to try it out since I keep burning tokens on pay-as-you-go on OpenAI and Anthropic.
ancient furnace
@humble nebula basic tool calls appear to be working for me but I haven't really run any extensive testing yet. Using:
"kimi-coding": {
"baseUrl": "https://api.kimi.com/coding",
"apiKey": "redacted",
"api": "anthropic-messages",
"models": [
{
"id": "k2.6-code-preview",
"name": "Kimi K2.6 Code Preview",
"contextWindow": 262144,
"input": ["text", "image"]
}
]
},
fleet totem
rapid warren
anyone managing/running vllm or sglang kimi 2.5/2.6 by any chance? openclaw wont do any tool calls with them. Let me know is you do work with them. thanks. (I mean running kimi on a server/system, not API from a provider)
rare tide
Kimi 2.6 feels horrible in openclaw (coding plan). especially compared to glm 5.1 - anyone else with that experience?
tired rampart
I am using minimax m2.7, would kimi k2.6 be smarter? I am using the 10 euro per month does kimi have similar subscription?
tired rampart
Kimi 2.6 feels horrible in openclaw (coding plan). especially compared to glm 5....
I have not used glm 5.1.... but I found 5.0 sooo bad, 1st very very very slow. and 2nd I had so many rate limit errors. I tried minimax m2.7 with the exact same setup and it worked so much quicker and did not have rate limits. But I prefere a working system than a non functioniong smart one.
unkempt vault
Kimi 2.6 feels horrible in openclaw (coding plan). especially compared to glm 5....
What's wrong with it? I've used GLM but man got rate limited last night after not using it for a while.
vestal notch
I can't complain. I had a coding problem today and after trying for an hour assisted by my local Qwen2.5-coder:14b I switched to Kimi 2.6 via OpenRouter and it nailed the bug in 60 seconds - and a second one that I have not even been aware yet. Got me charged $0,0194 for 17k tokens and it was like talking to a friend. Friendly, precise, not chatty, not over explaining. Just enough to point out the problem and suggest 3 options. I have seen others, namely GPT that tried to nag me for more and more prompts with "I could offer an even better solution than the current one if you are interested..." 4 times in a row ... I felt like in a telephone-track call (keep him talking as long as you can).
When Kimi2.5 was introduced some time ago, I had a lengthy chat with it and it felt very similar to the small brother of Claude Sonnet. It even claimed to BE "Claude" and we had a serious quarrel about that. He insisted that is insecure to use a Chinese based model and I should better trust Anthropic models "like him" for sensitive data ... On his suggestion we tried the "strawberry" test to determine the true identity - and the response was much more similar to "Claude Sonnet" than any to other model I tested. I don't think that's just a coincidence.
Certainly I would not use Kimi2.6 on OpenClaw running a MB task every other hour that burns 50k tokens on every turn just for fun. But for really useful work I give it a thumbs up for value/money.