we had some fun news to share but got side tracked we did our first test run on a new GPU rack with our new deepseek inference setup the price/perf we're getting is better than every single vendor that we've found. and there's still more optimization to do
177
58
30
4,378
264,672
Can you guys confirm you're not reducing the performance of the model at all though because I feel like it's a bit of a scam to say that you get X amount of usage when it's not the same model
1
151
31,624
any time we offer a model it's the exact same deployment as the official lab who made it the worst we do is serve at lower TPS to handle more volume but it is the same model, if we chop it up we release it under a different name

Oct 18, 2026 · 4:13 PM UTC

14
2
8
597
52,855
RelevantRecentLikes
Okay, this is very good to know because I've heard other people say that it is nerfed compared to the direct deepseek API. I will trust your word on this though!
1
11
3,962
Just re-engaged my subscription! I thought it was too good to be true to be honest, but if you guys are actually giving that much deepseek usage for $10 a month, this is a steal
2
3,141
Can't wait for the 1017 version!
2,518
Bruh, when DS V4.1 Flash gets upgraded to 1017?
3
71
10,588
team is working on it as fast as they can
18
2
5
397
29,003
A name like big-pickle maybe?
279
Are you sure? What about orcarouter, luminal, futurixai, novita, manhattan-project behind de same model id? What is it?
570
Good to know, thanks for being open about it
939
Isn't reducing tps the whole problem? Right now it seems no one can give me a decent tps anymore, except Google Vertex on API tokens.
1
1
700
In case anyone cares. I did validate this. Session headers are identical between the paid and free versions of models. It's just cut context + lower max output tokens to save on inference cost. Still insane we can get it for free.
2
608
make sense, computed must be enhanced
133
Less than a month between failing to guard access points. Such a failure
702
same model ≠ same system. 2s vs 56s TTFT changes timeouts, compaction, and turn count inside an agent loop. Same weights, different output — VulKan's −36% output tokens is likely that, not a different checkpoint. He already conceded "lower TPS," so this part isn't in dispute.
233
那你的 tps 太低了吧,DeepSeek 官方的跑到 200,你们就是 20,那我为什么不用官方的?如果有需要可以联系我,我可以帮你们订阅 kimi glm DeepSeek,只收非常少的费用,也许是一次 1 刀?
19