Awesome!
Thanks, that's perfect for my little STT tool (push to talk style)!
Even on CPU, 3x speedup! Now we are talking!
You are in the Acknowledgments! I didn't even have to add you. Qwen3.8-27B found the local benchmarks so impressive it was one of the first edit it made to the README after implementation 😎
let's gooooo! ;)
Guys ! After using it everyday since then, I can really say it's now been proven. Accuracy is better than the original parakeet-v3, I can feel it because priorly, parakeet-v3 has been daily driver for almost 6 months! Really good job just for this. And the speed is GAME CHANGER for my use case, really!! I'm enjoying using my own tool even more than ever!
So, thank you, and thank you again!!
Guys ! After using it everyday since then, I can really say it's now been proven. Accuracy is better than the original parakeet-v3, I can feel it because priorly, parakeet-v3 has been daily driver for almost 6 months! Really good job just for this. And the speed is GAME CHANGER for my use case, really!! I'm enjoying using my own tool even more than ever!
So, thank you, and thank you again!!
Running on Photon?
yup
Isn't it limited to only a few GPU models? Which GPU you're running it on?
you are right Luka, and I just realized I completely dropped support for the original parakeet model in my tool! 🤦 omg I hope the few users are still able to run it!
I forgot to check if it was running with my old laptop!
The story for myself was:
- I checked the compatible hardware
- As you also did, saw only few GPU were supported
- Seeing RTX PRO 6000 Blackwell, I still wanted to try with my RTX 4500 PRO Blackwell, thinking they might have just tested on the 6000 because that's what they had but it might work for me too
- Was working perfectly, so I ended up trying on my CPU (quite "old" AMD Ryzen 9 5950X (AM4 socket)) just to see
- Working perfectly too!
- Excited that even on CPU, gains were so high, I assumed ppl would only run it on CPU anyway cause there was no point running it on GPU for the tool specific use case
- Completely swapped older parakeet v3 to this Moondream's ultra one!
In the end, what I can say for now is that:
- their GPU requirements are, as they said, simply a "tested on" list, but some other GPUs support it just fine as well, like my 4500 PRO, it's not "hard requirements"
- at least my specific CPU arch, it's working perfectly, and I indeed have a MASSIVE speed gain, and accuracy only seems better than parakeet v3
I'll try with my old laptop both with CPU and GPU and report back to you soon (IIRC, CPU is an Intel i5 9th series, and GPU is a RTX 1660 Ti laptop). I keep you updated!
Alright so it's working on the laptop, both cpu and cuda. For some reason I get some NeMo warnings on cuda which are actually surprising cause NeMo shouldn't be used anymore since Photon entirely replaced it for this ultra model... I need to look into it. But otherwise, working without issue.
Thanks for sharing it! I was sceptical about giving it a try because of their limited GPU support.
You welcome friend! Yeah you should try! Oh and wasn't really clear: but the warnings I got are because of my code, not the Photon engine! They don't use NeMo at all.