@marcinteodoru: Kimi-K2.7 has released! latest coding model, is now available and open-sourced! While OpenAI and Claude code are battling, these open source models are quickly gaining traction and with it being six times faster and 10 times cheaper, it’s just a matter of time. #greenscreen

RanzJan :

I don't care fast. I care good answer.

mencuddling :

But how good is the coding? I’d take a slow but really amazing ai over a fast and sloppy one

farmsteadguy :

servers are always busy. not worth even trying

Cinder Works :

Tried it it’s pretty bad

Solomon :

Yes for 100 users 😂 then it will be garbage like DeepSeek

b_e_n_j_i_86 :

still only 256k context window, not 1.5m ...makes it hard to use over the likes of opus of even deepseek

tiktok6689765 :

first question "compare yourself to claude fable 5" it responds "I don't belive fable 5 is an actual anthropoc model ..." that's all I needed to see, thanks anyways

user99 :

Lame duck :

when we get fable 5 man? is there a chance by December as Christmas gift?

vibecoderscafe :

They distilled fable for a few days. Wish they had more time. 🥺

️zip :

i doubt the open source part

dgddxhhbb :

is it free ?

MSM Gaming :

I Use Kimi From Then It Was Made It's Insane.

Andrew D'Aurelius :

I already cut Codex usage by 3/4 by delegating the reasoning to a ChatGPT chat connected to github... Maybe this one can lift usage even more 🫡

Just Hit Send :

This is the part people miss with cheap coding models: API price per million tokens is only half the story. In agentic coding, tokens compound quickly because every debugging loop can resend project context, previous chat history, tool output, diffs and retries. So even if a model like DeepSeek or Kimi is cheap per token, a deep coding session can still burn through hundreds of thousands or millions of input tokens fast. That can remove the cost advantage compared with a flat monthly tool like Claude Code. Cheap models are great, but the real test is workflow efficiency: how much context they need, how many loops they take, and how often they get the fix right the first time.

Capitán Obvious :

so how are you running it for code? just as a CLI?

Руслан Фомичев :

just for info k2.7 scores ~29% on newest deepswe bench, opus 4.7 max 54%, opus 4.8 max 58%, gpt 5.5 xhigh 70%

C0deConnoisseur :

once you Kimi you don't go back.

биг боб :

the worst model

An :

i think this is just an ad

Dan :

US might need an open model to stay in the race.

I ain’t fucking shit :

Kimi is not really like codex and Claude.

Bat Nap :

Better than fable 5?

chunky fresh :

who should I invest in?

colmkeogh53 :

What is the cheapest video creator for Claude?