Abacus AI Dragon Models Are Coming To Open Source.
Finally, open-source is finally catching up to closed source
We are going to soon be launch a mini Flash model.
A good alternative for DeepSeek flash for mini agentic loops
Users are excited about Abacus AI's open-source Dragon and Smaug-Flash models for agentic tasks because small fast models enable cheap parallel agent loops and could let open-source win on cost, though one reply doubts the benchmarks.
No Digg Deeper questions have been answered for this story yet.
🚨 Open Source Abacus AI Dragon Models Are Coming
Open-source has almost caught to closed source.
We are actively working on
> creating a flash size model for small agentic loops > best for 1000s of parallel tasks
Name: Smaug-Flash
@bindureddy I almost thought you forgot about Smaug. When do we get Smaug Fat Dragon though?
@SkyBlueHarbor Wait for the new versions. Please use other models in the meantime
@bindureddy the fact that messages take up one credit or less for most messages is cool, but it doesn't keep context at all so that's a bummer.
https://huggingface.co/OpenTransformer/AGILLM-4.3 if you're thinking about making a Transformer Neural Network i got one here; i made it; it Transformer Neural Network Mixture of Expert Diffusion Blocks (let train layers to reduce vram or compute if needed). Also it can train as federation with internet network speed which is pretty cool. It includes Autoregressive generation/ Semi autoregressive generation and Non autoregressive generation which is really fast; and unlike diffusion inference which usually slow on cpu, it really fast on cpu too; tho gpu would faster of course.
@bindureddy Abacus AI this is super interesting
@bindureddy This is SICK! I think your platform is outstanding I use it everyday ALL Day LONG and it is so Fast and consistent cc @abacusai Been a very happy PRO TIER CUSTOMER since MAY 2025! SUPERCOMPUTER is insane!
@bindureddy 这名字一听就很会喷火
@bindureddy Anticipating 💥
@bindureddy almost caught up isnt the same as caught up tho
and naming your model after a dragon doesnt make the benchmarks real 😐
@bindureddy Good morning
@Semite2024 Fat Dragon on the horizon 🤣🤣
@bindureddy Flash model hype is real can’t wait to test it in my stack.
@bindureddy throughput's the easy win for fan-out. small models usually start drifting a few steps into a loop though. does smaug-flash hold reasoning depth across steps or is it more about per-call speed?
@bindureddy Breathtaking! The future of film making will be truly powered by AI
@bindureddy Cool
@bindureddy i do. i used up most the credits in a couple days just talking to fable 5 lol
@bindureddy 名字非常霸气
@bindureddy Cool!!
@bindureddy Small models purpose-built for agentic orchestration are the real scaling game, not just beefier tokens.