this is why you think open source models aren't good
@thdxr191K followers
288,338 views
47 Comments
Comments from the original post on X (Twitter)

@kristoph
this is a common open coder issue; especially if you run theough groq. kimi behaves much better using their api or if you host it yourself
29 likes

@ibuildthecloud
This reenforces that fact that I don't care what anyone says about model performance. There's so many layers on top that people don't understand and rightly or wrongly attribute to the model.
22 likes

@dubaplex
I am going to use opencode zen for my apps inference and there is nothing you can do to stop me
18 likes

@thdxr
@realduba we WANT you to do this the more usage we get flowing through the cheaper rates we can get for all of us use it and abuse it
67 likes

@jayair
@realduba kinky
18 likes

@dubaplex
@thdxr 😏
1 like

@_SpaceKangaroo_
what are openrouter and groq specifically doing wrong to cause this? I get latency but if the prompt, model, and settings are all the same how? The model itself must be different somehow if this post is real.
9 likes

@YayaSoumah
So do you agree now that implementation and vertical integration matters? A few weeks ago you was basically saying that models behave all the same in each application and that there was no such thing as model optimization
9 likes

@thdxr
huh - this is a video of the exact same application side by side opencode zen is a provider, it'll be good in every application you use it with assuming they didn't fuck something up
19 likes

@YayaSoumah
Then let me reframe, if the model is the same, why Opencode zed version behave much much better?
3 likes

@thdxr
the idea here is the model is not the same when 2 providers claim to say "we provide kimi k2" you think you're getting the same product but they can serve it in ways that are good and in ways that are bad we just did the work to try out all the different ways to host it and only use the good ones
59 likes

@YayaSoumah
Love this! Keep up!
2 likes

@berenddeboer
Yeah, this was a significant change, in the past I found opencode frustrating as I tried all these new models, but they were basically unusable. That's just fixed now.
8 likes

@BrandGrowthOS
Provider infrastructure makes all the difference. Seen this with Claude too - same model, wildly different performance depending on who's serving it. The tooling layer matters more than people think.
8 likes

@AdamPippert
It’s almost as if the model is not the final product you receive from the API… Good job with zen, it’s clear you care about the final presentation and not just the model itself.
7 likes

@jayair
This is actually kinda crazy
7 likes

@olvrgln
genuinely curious what's going on here. what's the secret sauce, is it literally just default parameters like temperature, etc?
6 likes

@molerburb
i honestly think u should make the list public, so the bad providers are forced to correct it. though i also understand it's hard work and you may wanna keep it close to capture value
5 likes

@thdxr
nah we will make it public (it's already visible in our github repo) - we just haven't been loud about it yet because we just launched this service and things may shift around
13 likes

@molerburb
very nice to hear that!! appreciate it a lot!
























