MoonMath.ai

@ItayAmir8 @MokadyRon @bria_ai_ You are the 👑

@MokadyRon @bria_ai_ It was great working with you guys @moonmathai !!

Making models run fast at inference requires optimizing the entire AI stack. It was great partnering with MoonMath to take @bria_ai_ 's Fibo to the next level of speed. Unlike standard models, Fibo consists of a Reasoner (VLM) and a Renderer (Flow Matching), requiring both to be optimized at the algorithm, deployment, and kernel levels. And most importantly it was great to work with @moonmathai Read more in the new blog post

We accelerated Fibo on Hopper, Instinct & DGX Spark (Blackwell)! Awesome model and awesome team.

@OPatashnik @DecartAI @ingeration @Lightricks Explore talks from last event: youtube.com/playlist?list=…

📢Announcing the 2nd Generative Media TLV meetup, April 15th. Tel Aviv. Speakers: @OPatashnik Tel Aviv Uni Essam Qeys , @DecartAI @ingeration , @Lightricks no fluff ! Sign up: luma.com/pnrjjkyf

“24 FPS” ≠ real-time Example (seaweed-apt.com/2): - 24 FPS (this is throughput!) - ~160ms latency → ~4 frames delay That’s not interactive! FrameCommit moves latent video models in that direction. True frame-by-frame real-time a-la @DecartAI

8/ A final pixel-mismatch loss penalizes decoded outputs that disagree with already committed frames, reducing jitter. The proposal is simple: keep the latent video model, but change the conditioning/inference loop so it behaves like a 1-pixel-frame-per-step live-stream system.

7/ Training starts from a StreamDiffusionV2 / CausVid checkpoint. The new cross-attention layers are randomly initialized, and α is annealed from 1 -> 0.5 so the model gradually learns to use committed-frame conditioning before standard fine-tuning.

1/ New post: FrameCommit: Journey From Wan to Decart LSD, Part1 The target for live-stream video is not just high FPS but ~40 ms input-to-output latency per visible pixel frame. moonmath.ai/posts/framecom…

The Quantization Handbook 🪄 The first SoK for diffusion quantization 📘Read: moonmath.ai/assets/posts/q…

This week we lost Chuck Norris, or so reality claims. In his honor, here are 10 facts about Chuck Norris and AI. Feel free to add your own 1. Chuck Norris won all reward models 2. AI trains on Nvidia because it cannot keep up training with Chuck Norris 3. When a model looks at Chuck Norris, it backpropagates 4. Chuck Norris defeated AlphaGo in Go 5. RLHF was invented by Chuck Norris, feedback originally was his kick 6. Your AI agent will go to sleep before Chuck Norris 7. Chuck Norris models don't need training, they have zero loss 8. Chuck Norris can get as many H100 as he want 9. AI never hallucinates when speaking with Chuck Norris, it’s just sometimes afraid to tell the truth 10. Chuck Norris can compile flashattention so quickly, nvcc asks him for advice h/t my team, Thank you Chuck, we'll always remember♥️

had to integrate Recraft the moment v4 dropped. more coming soon 👀

🧑‍🏭 LiteRunner 🧑‍🏭 MLOps-Style Tracking Without Touching the Code (New Tool) TL;DR: LiteRunner adds lightweight tracking to any CLI command without changing the model, saving params, outputs, and metrics locally and in W&B so every run stays reproducible and organized. Code (open source!): github.com/moonmath-ai/Li… Blog: moonmath.ai/posts/literunn… Contributions are welcome 🙌 More background: When running video generation experiments with diffusion models, the workflow quickly turns into bookkeeping. Every run starts with hand-editing long CLI commands, quoting paths, swapping flags manually, and each run produces a different combination of config, output videos, metrics, and debug data. Output files end up scattered across multiple folders and machines with no central record, sometimes even overwriting each other. Moving those files and recording runs becomes tedious, and inevitably the one run that wasn’t properly recorded turns out to be the one that matters. Revisiting an old experiment often means digging through notes just to figure out whether it used seed 10 or 42. When you own the code, you can wire in an MLOps tool to solve this. But often you’re just a user of someone else’s model, and modifying their source just to get proper tracking isn’t practical. That’s when the idea comes up: instead of changing the model code, bring MLOps-style logging to arbitrary CLI commands, so experiments can be tracked without touching the original implementation.