used it yesterday in kimi code, it's great

They spent some time on design and motion design. +1

Tried 2.5 too? I am considering to grab one of their subs for some testing. o.o

Oops I guess I was using 2.5

I'm really amazed. it's solid and reliable, does not overengineer, straight to the point. And yes the frontend design is top notch. I like this AI lab, they have taste discovered kimi code too, very minimalist, refreshing TUI, similar to Pi

is this one of the "AI world clock" models that actually almost always makes a correct clock?

Where are you all running Kimi2.5 at? Locally? The only AI platform I use is MapleAI and they currently only have Kimi2. I have been liking Qwen3 Code. I try to run some on my laptop, but a little slow.i just mainly play with this all to learn.

ah yes, give him a refactoring task, or anything that would need heavy indexing and deep understanding of your codebase, this is where it shines the most

how does it rank next to Claude Opus 4.5 and Claude Code?

so the agent has better memory management? as in AI context

my referance is amp. they use claude opus and different model for different task. it's unmatched for frontend and design, give it a large task it will go through it super fast without melting your context and come back with very precise insights. It's producing very clean code and has strong problem solving capability. I haven't had much time with it but so far it did no mistakes, which is impressive. I would definitely use it as my workhorse, 99% of the time. best oss model, and for me best model overall, great taste for coding and design.

i got the answer from the horses mouth: The consistency you're seeing with the clock test reflects a few specific strengths in how I approach code generation tasks—particularly those involving state management, mathematical precision, and real-time rendering: 1. Robust Handling of Time-State Logic A working clock requires correctly calculating angles from Date() objects (hours × 30° + minutes × 0.5°, minutes × 6°, etc.) and managing continuous updates via requestAnimationFrame or setInterval. I tend to generate this logic with fewer off-by-one errors or rounding issues that break synchronization. 2. Spatial Reasoning + DOM Precision Clocks need CSS transforms (rotation origins, absolute positioning of hands) coordinated with the mathematical logic. I maintain consistency between the calculated angles and the visual representation—where some models generate the math correctly but botch the pivot points or layering. 3. Self-Correction During Generation When I generate a clock, I'm effectively "simulating" the execution in context: "If seconds = 45, the hand should point at 270°... does my CSS rotate from the bottom center?" This reduces the drift or misalignment issues that make clocks tick incorrectly or start at wrong positions. 4. Clean Separation of Concerns I typically structure the generated code with clear separation between: Time calculation logic DOM update functions Styling/animation Initialization/cleanup This architectural clarity prevents the "spaghetti code" that causes clocks to stutter or drift in less consistent generators. Why This Matters for Your Engineering Work These same capabilities translate directly to engineering applications: real-time sensor dashboards, structural analysis visualization, parametric CAD scripts, or simulation loops. The clock is essentially a minimal viable test of: "Can this AI reliably generate code that maintains accurate state over time while rendering geometric transformations correctly?" If you're working on computational design, BIM automation, or analysis tools, this precision with geometry + state management is exactly what you'd leverage for things like real-time load visualization, animation of structural behavior, or interactive parametric models. What kind of engineering visualization or automation are you building? I can show you how these same patterns apply to your specific domain.

I am using their one week free preview:

Kimi - 会推理解析，能深度思考的AI助手

not openrouter or any proxy

Kimi - 会推理解析，能深度思考的AI助手

the thing that is bothering me is the gulf between the middle option and the top option. seems a bit crappy that i can't have something in between the $31/month and $159/month plan tho. i'm kinda skeptical that it's gonna do as well for the money, because i immediately got the impression from first tinker with the freebie that it was slow. i mean, if it does the work, sure but... idk, i might try the $31 at some point soon, for now i'm in the groove with claude.

yeah, it's a linear increase in price, there is no premium in token/dollar for each subs. if you are not into ralph loop the medium plan might be enough. if you are a looper, well, the big plan get you covered

gonna trial the smallest one, anyway. they give you a week free. i can't say i wasn't a bit put off by how vigorously their web app pastes identifiers in the address bar tho. but meh, whatever, it's all the same really i think. i may just have not paid so close attention in the past when i poke at offerings on the web how they run their site. i mean, obviously, kimi is chinese product. but considering how practical chinese are, even if they aren't genius at engineering, they certainly are genius at reverse engineering, which is a kind of engineering that i groove on. when i see a black box, i want to penetrate it. no, not like that lol.