Okay, here’s my blog post about messing around with Llama 3 and 3.1:
So, I’ve been itching to get my hands dirty with the new Llama 3 models, and I finally carved out some time to do just that. I also stumbled upon this “3.1” thing, which seemed like a minor update, but hey, I’m all about squeezing every last drop of performance out of these things.
Getting Started
>
First things first, I needed to get the models. I grabbed the 8B and 70B versions of Llama 3, just to have a good range to play with. The 8B is nice and speedy for quick tests, while the 70B is the big gun for when you want to see what it can really do. And, of course, I snagged the “3.1” versions too. I wasn’t really sure what to search up, I just got on my computer and got them.
The Setup
>
I’m running all this on my home machine, which is a pretty beefy rig but nothing too crazy. I use the program where all the AI’s that I have saved are accessible. My process to test these language models out is all over the place, but I needed a solid way to put these models through their paces.
Initial Impressions
My first thought, running the Llama 3 8B, was “Okay, this is snappy!” It’s noticeably faster than the older Llama 2 models I’ve used, which is a huge plus when you’re iterating and just trying things out. The responses felt pretty coherent, too, which is always a good start.
Then I fired up the 70B model. Whoa. This thing is a beast. Obviously, it’s slower than the 8B, but the quality of the responses… yeah, there’s a definite jump there. It’s like the difference between talking to a smart kid and a seasoned expert. The 70B just gets it, you know?
Diving into 3.1
Now, for the “3.1” part of this whole experiment. Honestly, I wasn’t expecting much. I figured it would be some tiny tweaks under the hood, maybe a bit of bug fixing. And… well, that’s kind of what it felt like. I ran some of the same prompts through both the 3.0 and 3.1 versions of the 8B and 70B models, and I had to squint to see any differences.
- Slightly Better Coherence? Maybe? It’s hard to say for sure, but on a couple of longer, more complex prompts, the 3.1 versions seemed to stay on track a little better. But it was subtle.
- Speed? Nope, didn’t notice any significant speed improvements. Again, this is just my anecdotal experience, not a rigorous benchmark.
The Verdict (For Now)
So, here’s the deal. Llama 3 is a definite step up from Llama 2. The 8B is great for quick experiments, and the 70B is a powerhouse for when you need top-tier quality. As for the 3.1 update? It’s… fine. I wouldn’t rush out to upgrade if you’re already happy with 3.0, but if you’re starting fresh, might as well grab the latest version.
This is all just my initial tinkering, though. I’m planning on doing some more in-depth testing, especially with different types of prompts and tasks. But for now, I’m pretty impressed with what Llama 3 (and 3.1, I guess) can do.