https://kotlinlang.org logo
Join Slack
Powered by
Not a single non US-tech option. JetBrains, I'm d...
# random
d
Not a single non US-tech option. JetBrains, I'm disappointed.
h
You're complaining about US-only tech options ... with a macOS screenshot? 🤔
🤣 9
🙄 1
f
I think you can use a local Ollama instance, to run your favorite Chinese models like Qwen
➕ 1
d
I think he meant Mistral, Le Chat
👍 1
Jetbrains itself is non-US, it's European, I think it's enough
a
maybe its down the list? cause neither lechat nor qwen are in the top, in the charts
d
I agree, not top
s
I tried a lot of alternative and cheaper (even free models through OpenCode), but to be honest only Claude Opus and Fable are really usable for agentic coding. The lower models introduce so many subtle bugs that take so much time in debugging that I'm faster writing the code on my own. Especially MiniMax made a first good impression, but after hunting down so many subtle bugs and letting Claude Fable do a deeper review, I‘d say it’s unusable. When I review AI generated code it's pretty much like reviewing someone's else’s code. Yes, I catch a lot of things, but I don't think it through like you would when writing it on my own. I don't have those „oh wait, now I have a better idea“ moments when I'm generating code.
I question the pre selection of Gemini, too. That model isn't a good coder in my perspective.
So, yeah, they could offer non-US models, but I doubt that's useful when even most of the aus models are not suited for the job. :/
a
@hho what a strange take. Fine you can pick Linux. But according to you I should just accept all software should be US based when I use a Mac?
@Stefan Oltmann I have had some pretty good results with pi.dev icm OpenRouter and Kimi K3. But even Kimi K2.7 code works pretty well.
s
Didn't try Kimi K3 so far, but I'm pretty disappointed by the cheaper options. 😕
I'm currently trying to accept that I might just have to pay for Opus.
a
I used Claude Max since December with Opus. But I havent touched it since beginning of July with pi and OpenRouter. Although it is heavily project and task dependent. But pretty happy with the results and the speed. Can’t hurt to compare notes maybe.
s
Yeah, I'm right now reading that Kimi K3 can beat Opus. I try to take such news with a grain of salt.
In any case, it's very good to see Chinese models competing in this space. They show that things can be done more efficiently, and I also think they help keep prices low. If OpenAI and Anthropic were the only serious providers of coding agents, we would probably have already seen massive price hikes.
a
I take all marketing with a grain of salt. Opus is great but sometimes borderline stupid. Fable is dead slow and somehow overcomplicates everything and is horrible expensive. The proof is in the eating as we say. For me I am happy with open weights models like Kimi because it means that we might break free of the US dominance because they have become quite hostile. And so far Kimi K3 and even 2.7 Code are similar to me as Opus. And I have used Opus extensively the last half year. I like the OpenRouter model comparisons. It gives some great insights. Fusion models are also quite interesting
s
Today I switched back to Sonnet 5 to save on my quota, and it immediately introduced a subtle bug... So annyoing.
For me I am happy with open weights models like Kimi because it means that we might break free of the US dominance
I'd prefer open weights, too. Open source even more. For now I'm looking for something with decent results. Your experience with Kimi K3 motivates me to take a look. What does it cost you per month?
a
I am on API usage via OpenRouter. So depends heavily on the task. I normally use K3 for the design plan with separate tasks. Then 2.7Code for actual implementation steps and tests. I spend around 50 to 75 dollars per month. Claude Subscription is heavily subsidized by Anthropic of course so hard to compare sometimes.
👍 1
s
Don't use MiMo V2.5 - it's so hacky and dirty. It removed source code comments and replaced existing Composables like my 
IconButton
with something it build itself. That was my last attempt using cheap models.
a
Yeah makes sense. You pay peanuts and get monkeys ;)
💯 2
I did a nice experiment where I created a spec document to create a durable actor library (virtual actors) in Kotlin with Postgres as backing provider. I then created 2 new repos and in one let Claude Opus 5 do the work and in the other Kimi K3 (via openRouter). Results were similar and of similar quality as far as I could see. Same library choices, same Gradle setup, same latest versions of all libraries. Claude Opus total costs $28, but part of my subscription. Kimi K3 started initially very cheap, but lost the plot somewhere while debugging timing issues. Eventually it got solved, but for the 98% it was at $5-ish but ended at $20.... and thats pure token costs via OpenRouter. Claude ran via Claude Code CLI obviously, Kimi via Pi.Dev. I can imagine that harnassing can differ here quite a lot. I then let Fable audit them both for correctness and create a test harnass and they both had 1 critical issue, both in a completely different area 😄. Good learnings, and running the actor framework now on my own server for doing cronjobs and long running things.
🚀 2
👍 1
s
That's how I spent my free $100 ... doing a deep architecture & code review of everything Codex, MiniMax, MiMo, Sonnet and so on did.
I think discussing a plan and fleshing out the details with a higher model and just have a lower model generating the code (while strictly following the written-down plan as TODO.md) works best for me.
💯 2
Ah, finally I found some numbers on Opus 5 vs Kimi K3. quesma.com/blog/baba-kimi-k3-opus-5
> Kimi via Pi.Dev Not OpenCode? Or JetBrains AI Assistant? Oh wow... so many options.
a
So many options. I am not always working in JB products and have never tried OpenCode ;) i guess pi and opencode are similar
3 Views
Open in Slack
PreviousNext
Powered by