Introducing RHFM, our custom mega kernel that we are using to redesign intelligence for the edge, inspired by @flappyairplanes, @HazyResearch, and @cohere. We built a persistent GPU runtime that trains a model, produces predictions, and plays Doom, all at once.
1/ 🧵
RHFM keeps one CUDA runtime running. Instead of launching a new kernel for every operation, we send commands to a resident kernel that holds the execution roles needed for language training and Doom. We go into more technical details on our blog: axioniclabs.ai/research/rhfm…
2/
We believe that this kind of work will enable more cool experiments in deep learning, particularly when it comes to alternative solutions to the credit assignment problem (replacing backprop) or zeroth order optimizers, such as those pioneered by @@industriaalist and others at QLabs.
Because our team does not sleep and has no life (a lot of supplements and caffeine), we decided to work on a mega kernel for ResonatorLM, inspired by the recent demo @flappyairplanes.
A 500M parameter model learning live while it plays DOOM, all on one persistent kernel.
Check out the full video here, all implemented on a single @nvidia RTX PRO 5000 Blackwell card.
A 4B parameter model with strong agentic and long context capabilities, running in your pocket, on an @Apple iPhone.
axioniclabs retweeted
Our Resonator model from @axioniclabs now works locally on a iPhone.
A 4B parameter model, that is running completely locally, and has agentic intelligence + long context capabilities.
A lot more to come soon!
A big shout out to the research team for getting this out!
Here is the official launch video! A shout out to @typesafeai and the creators of Anything-Jev for inspiring this work!
We built a system 1, Jev-like decision model in 2 days last week based off our own Resonator architecture. It is significantly faster and more efficient for specific use-cases. It works surprisingly well for specific classification and agentic use-cases:
Read the full post, and more what we actually do, here: axioniclabs.ai/research/deci…
We are continuing to scale up Resonator.
We have a few results that are dropping later this week, centered around our pretraining pipeline, and an updated 4B parameter model.
We presented ResonatorLM at ICANN this afternoon.
We think long-context language modeling and continual learning need a different architecture, and that is the problem ResonatorLM is working on.
If that is the kind of work you care about, our blog is at axioniclabs.ai/research/reso…
axioniclabs retweeted
Our account was temporarily compromised, We will NEVER launch a token!!!!
There is no Axionic Labs token or crypto affiliation. We will never launch a token. Any and all tokens launched under the Axionic name are fake and a scam. Our Twitter was temporarily hacked. Please be careful!
@axioniclabs research is being published at ICANN 2026. Our research on post-transformer alternatives will be highlighted in the main conference from 14–17 September at ICANN, held at University of Padua
ResonatorLM explores an alternative to self-attention for long-context sequence modelling, using learned resonators to build a more efficient recurrent architecture.
At @axioniclabs, we work on new model architectures and learning systems that challenge the assumptions behind today's AI stack.
ICANN brings together researchers working on the foundations of neural networks. That is the hardest audience this work could have, which is exactly why we want it in front of them.
Come talk to us if you're working on post-transformer architectures, continual learning or adaptive AI models.
"Axionic Labs is not hiring."
That's what we've been replying to DMs all week ; a lousy thing to send someone who read our work.
So: the Resonance Prize. Read our research, tear it apart.
$500 for 1st
No ML background needed
Entries close Sept 21
Rules ↓
axioniclabs retweeted
Today, we raised a $13M series A to solve the GPU shortage
Trillions are being spent on GPU CapEx while 80% of it sits idle
The capacity already exists
Virtualizing GPUs unlocks it
Backed by Matrix, YC, and CEAS
We recently announced a new paper, centered around a post-transformer architecture designed for language modeling. A 🧵follows:
We are not the first to explore this architecture: independent researchers have proposed a physics-based alternative to traditional transformers in forums such as HuggingFace and Reddit. However, we are the first to formalize it.