LFM3-VL returns its first token on an image question in 200 ms on Qualcomm’s Dragonwing IQ9.
The fastest llama.cpp setup we tested on the same board took 710 ms.
The video shows LFM3-Omni-2B on the same board. How we did it: zetic.ai/blog/lfm3-on-dragon…
We build tailored inference engines for your model on your target hardware: zetic.ai/#contact
Partner program for the ixian app,
Join the waitlist: forms.gle/6KnNYAHvKucBPS1v6
#PhysicalAI #Robotics #AIInference #OnDeviceAI
heath retweeted
Hiring an inference optimization engineer for your robot?
Meet ixian agent. Give it your model, board, and targets, and it builds the engine.
Our team vs PyTorch eager:
π0.5 on Thor: 9.7× faster
d1-2B on Orin: 4.9× faster
Design partners wanted. DM me your model + board.
heath retweeted
d1.2-ultra is the #1 model on the Artificial Analysis Decision Index v2.1, with d1.2 placing #6
@liquidai has released d1.2-ultra and d1.2, updating their d1-ultra and d1.1 flagship decision models.
Key results:
➤ d1.2-ultra is the #1 model on AA-Paperclip. With a score of 76, d1.2-ultra is the leading model on AA-Paperclip, our first-party private benchmark that assesses decision-making and extraction capabilities. It leads the next-best model, Drex Max v1.1, by 3 points
➤ d1.2-ultra and d1.2 improve over their predecessors. d1.2-ultra scores 75, a 4-point gain over its predecessor d1-ultra, while d1.2 scores 69, a 3-point improvement over d1.1
➤ d1.2-ultra and d1.2 both sit on the Capability vs. Cost Pareto frontier. d1.2-ultra and d1.2 cost $133 and $63, respectively, per one million Artificial Analysis Decision Index tasks and both models sit on the Capability vs. Cost Pareto frontier
Additional model details:
➤ Context window: 64K tokens
➤ Pricing: $0.042 and $0.02 per 1M input tokens for d1.2-ultra and d1.2, respectively
➤ Multimodality: Text and image input
➤ Providers: Available through Liquid AI’s first-party API
heath retweeted
We're open sourcing WendyOS, a physical AI operating system for NVIDIA Jetsons, Qualcomm Dragonwing, Raspberry Pis that allow you to deploy with just a USB-C cable.
github.com/wendylabsinc/Wend…
heath retweeted
Great evening yesterday at @liquidai !
@vivmarquez & @heath0xFF hosted the Liquid AI x @Qualcomm Dragonwing iQ9 hackathon and the winning team got an iQ9 board PER TEAM MEMBER !!
@zeticai_ provided the tailored inference engine to run the LFM3 and d1 models on iQ9 with the best performance in the market.
Should’ve gone for the price instead! @zetic_ai
heath retweeted
This is NOT your usual AI event 👀
Next Wednesday, @liquidai is teaming up with @AISocraticOrg for a very different kind of evening about on-device AI agents.
@feulf has been carefully curating the guest list, bringing together engineers, researchers, and builders for a Socratic evening where everyone gets to participate, challenge ideas, ask questions, and contribute to the conversation.
What to expect?
🌶️ Spicy questions (can a 1B model really compete with a 70B model?!)
🧠 Deep technical discussions about on-device AI
🎤 Insights from some amazing speakers: @mlech26l, @zetic_ai, @alanzhuly, @Henry_Ndubuaku
😂 Plenty of memes
🚫 ZERO boring corporate slides
🤝 And a room full of people who are just as excited about this stuff as you are!
The idea is simple: less passive listening, more actual conversation. We want everyone to leave having learned something from each other.
📅 Wednesday, November 18
Spots are limited to keep the discussions interactive. Apply to attend! 👇
luma.com/on-device-agents?ut…
heath retweeted
Classic machine learning is BACK, BABY
heath retweeted
Friday 11/13
‣ On-Device AI Hackathon: Cleared for Takeoff luma.com/0nsdnkoa
@liquidai @vivmarquez @qualcomm @heath0xFF
‣ 8090 AI Hackathon 8090-solutions.link/x-lists-…
@AfifaSiddiqua_ @sojoodi @carinaproduct @ujustgotleid
heath retweeted
Replying to @heath0xFF
Coding is saturated. What's still very hard is implementing physical AI in industry.
heath retweeted
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
huggingface.co/spaces/huggin…
heath retweeted
ExecuTorch on Qualcomm: Train, Optimize, and Deploy to the NPU teapaw.org/i/broadcasts/2nQBYnmVL…
Liquid 🤝 @Modular ✨
LFM3 from @liquidai is now running on MAX. 🧑🚀
All 6 models in the suite are supported, across text, audio, vision, and omni. The implementations are open source, so you can read the code, run it, and build on it: max.modular.com/api/python/p…
Which modality are you trying first: text, audio, vision, or omni?
heath retweeted
Standard security cameras are useless - they just record people stealing your stuff. So I teamed up with @Qualcomm_Dev to build an AI camera system that screams personalized warnings at intruders instead. 🔈👀
Starlink Mobile is building the first fully-integrated US nationwide satellite and terrestrial network that will keep people connected on their phones no matter where they are. Our 5G base stations will leverage @Qualcomm chips to accelerate deployment, starting in 2027 🛰️📱
SpaceX has selected Qualcomm #Dragonwing Cellular Infrastructure solutions to support @Starlink Mobile’s nationwide terrestrial 5G network in the U.S.
heath retweeted
SpaceX selected Qualcomm Dragonwing to accelerate @Starlink Mobile’s nationwide terrestrial 5G network. Together with @elonmusk and the @SpaceX team, we’re excited to build world-leading cellular infrastructure and a path to AI-native 6G. 🤜🤛
SpaceX has selected Qualcomm #Dragonwing Cellular Infrastructure solutions to support @Starlink Mobile’s nationwide terrestrial 5G network in the U.S.
heath retweeted
Please just ask Astra to use Antidoom instead, it's 100x cheaper and faster.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
heath retweeted
Awesome work by @_JoshuaTimothy! 🚀
Great to see @Qualcomm GenieX powering real-world agentic AI apps with @liquidai LFM3 on @Snapdragon NPUs! GenieX also is used by academic for on-device AI benchmark now, see this paper: arxiv.org/abs/2610.40039
I'm building the coolest agentic mobile app on @Snapdragon that analyzes and sorts all of your videos and photos on device! This all runs locally on the 8 Elite Gen 5 NPU using @Qualcomm_Dev GenieX and the new @liquidai LFM3 models! Video editing just got a whole lot easier!
heath retweeted
⚖️ Nov 3 - Decision Models SF Demo Night
We're kicking things off with @maximelabonne, who'll tell us everything about decision models!
Open call to builders! If you've been building anything with our decision models (even a scrappy prototype), we'd love to have you demo it to the community!
👉 RSVP: luma.com/ejl2hloq?utm_source…
👉 Open call to present: forms.gle/GwHePAVx6QQxNEra8










































