Skip to main content

Mo Moustafa

AI product engineer and creator of Olly (olly.bot), an AI assistant in iMessage with 250K+ users; guest on Packet Pushers' Day Two DevOps discussing OpenRouter and open-source models.

Nov 4, 2026

13:27
What behaviors are you seeing that are different and how do you control for that?
13:32
Yeah, that's a great question.
13:34
The main thing to know is each provider gets the raw weights of the open source model and And then each of them has different hardware.
13:54
And that is the first layer of differences.
13:58
The hardware might mean that one of them has more RAM compared to, let's say, like flops available, or some of them have more flops than RAM.
14:12
And so in order to make the to make the best use of their hardware because they don't want to to waste any um any of the capacity that they have by not utilizing all of their flops or all of the RAM, they might make some different parameter optimizations.
14:33
So each model has a bunch of parameters.

6 MINS LATER

20:33
story.
13:27
What behaviors are you seeing that are different and how do you control for that?
13:32
Yeah, that's a great question.
13:34
The main thing to know is each provider gets the raw weights of the open source model and And then each of them has different hardware.
13:54
And that is the first layer of differences.
13:58
The hardware might mean that one of them has more RAM compared to, let's say, like flops available, or some of them have more flops than RAM.
14:12
And so in order to make the to make the best use of their hardware because they don't want to to waste any um any of the capacity that they have by not utilizing all of their flops or all of the RAM, they might make some different parameter optimizations.
14:33
So each model has a bunch of parameters.

6 MINS LATER

20:33
story.
13:27
What behaviors are you seeing that are different and how do you control for that?
13:32
Yeah, that's a great question.
13:34
The main thing to know is each provider gets the raw weights of the open source model and And then each of them has different hardware.
13:54
And that is the first layer of differences.
13:58
The hardware might mean that one of them has more RAM compared to, let's say, like flops available, or some of them have more flops than RAM.
14:12
And so in order to make the to make the best use of their hardware because they don't want to to waste any um any of the capacity that they have by not utilizing all of their flops or all of the RAM, they might make some different parameter optimizations.
14:33
So each model has a bunch of parameters.

6 MINS LATER

20:33
story.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.