CodecFlow

Robots are ready for the real world, but the infrastructure behind them isn’t. CodecFlow is building the execution layer designed to close that gap. Watch the full breakdown and see what that future looks like.

Vercel didn't beat AWS. It made AWS feel like git push. SimArena does that for robotics simulation.

Most robotics companies raised VC and built SaaS. We fair launched on @Pumpfun, won their hackathon, and built a community at the intersection of robotics and crypto that believed early. This is one of our edges, and a distribution model few can tap into.

@peaq Liquid machines era begins. Excited to see peaq cook.

Today we’re launching a whole new asset class: liquid machines peaqOS is finally here and it’s ready to turn robots and machines across chains into investable assets and financial actors It’s time for blockchain to live up to its full potential app.peaq.xyz/os

Anyone with an idea for a robot should be able to simulate it in a browser tab, train it, and deploy it from one place. That's the bar SimArena is aiming at.

@PopPunkOnChain @Pumpfun how do i triple like this?

This is a common shortsighted take that I expect to disappear soon. Obvious disclaimer first: I'm not endorsing any specific project/token. I'm speaking from my personal experience and observations. Your mileage may vary. You don't launch a token on @Pumpfun as a builder because you don't have the technical know-how to launch a token. You launch a token on Pump Fun because of the immediate distribution layer it gives you access to. Having a token tied to your project can provide you with a pretty strong signal for how much people truly care about your product. Are people excited about your product, talking about it, and actively trading the token? Might be a strong indicator that you're building something useful. You can use this to determine the level of interest potential users, and then VCs will have in your idea. The trading activity generates creator fees for you. These can be used to bootstrap the early lifecycle of your product. Then, you can integrate your token as a core part of your product, like we are currently doing with @pumpcade. This is exactly what I did with Pumpcade (and many other projects did as well). The creator fees sustained me quitting my Senior Protocol Engineer job and building the MVP of my product. I know how to launch a token. But I CHOSE to do it on Pump for the above reasons. This later allowed me to win their hackathon, raise 6 million dollars across 2 back to back oversubscribed funding rounds from top tier VCs, and now scale my team. Like I said above, your mileage may vary. You should always do due diligence into the builder behind the token, but a token on Pump is one of the strongest bootstrapping mechanisms for a company in 2026. I've raised venture capital without a token and i've now raised it while having a Pump Fun token. I'd pick the Pump Fun token way every single time. We will see unicorns built on Pump Fun.

Physical AI is coming. We're building the playground.

Open source won web infrastructure. It'll win robotics too. Look at the pattern. Web servers: Apache, then Nginx, then the entire cloud-native stack (Kubernetes, Docker, Terraform). All open source. Proprietary web servers are a niche now. Databases: MySQL, PostgreSQL, the explosion of open-source data tools. Even AWS’s biggest services are built around managed versions of open-source databases. AI/ML: PyTorch became the default for research and much of production. Hugging Face became the model hub. LangChain, LlamaIndex, vLLM, all open source. Even when proprietary models lead, the surrounding ecosystem catches up fast. Now look at robotics. AI and robotics have been advancing faster than almost any field, largely because of open research and shared tooling. Papers, repos, and frameworks build on each other in real time. Tools like ROS2, Isaac Sim, and Dora are pushing the space forward. But they’re fragmented, heavy, and not designed as a unified stack. ROS2 handles middleware well, but not cloud or multi-agent coordination. Isaac Sim is powerful but closed. Dora is promising but still early. The result: every team builds their own integration layer. Nothing is reusable. Nothing compounds. CodecFlow is open source because we think the winning infrastructure stack will be open. Not because of ideology, because of economics. Open source creates community contributions that improve the product faster than any team could alone. Trust, because teams can audit, modify, and extend the stack. An ecosystem of plugins, integrations, and tools built by the community. And talent, because developers want to work on and with open-source tools. The proprietary robotics tools of today will be the legacy systems of tomorrow. The open stack will be the default. We're building that stack.

Procedural Assets, now in SimArena. Most simulators rely on static 3D models. You import them and place them where you want. SimArena takes a different approach with procedural assets defined in code. The difference: • Standard approach: Import model → Place → Done • Procedural approach(SimArena): Define parameters → Generate → Customize as needed Benefits: • Quick scenario testing • No need to search for specific 3D models • Easy parameter changes • More training variety from the same asset

@codecintern @Pumpfun Our intern❤️

@codecopenflow @Pumpfun my favorite project of all time is now 1 year old. eyes so blurry i can’t even type. $codec coded.

Today marks one year since we launched on @Pumpfun. We started as on-demand cloud desktops for AI agents, powered by MCP and TEE. The idea was simple: give agents real computers to work with, not just browser tabs. Windows, macOS, Linux. Machines that could run anything a human could. So we built more. Fabric: a cross-cloud orchestrator that finds the right hardware in real-time across centralized and decentralized providers. optr: a toolkit for building operators that see, reason, and act across desktops, browsers, simulations, and real robots. RoboMove: proof that the stack works. SimArena: robotics simulation without the expensive GPU barrier. We expanded into embodied AI and robotics because that's where the real problems were. When we launched, we found a community that actually wanted this. People showed up, asked hard questions, tested our early demos, and stayed. That community pushed us to think bigger. We won the Pumpfun Build in Public hackathon because we never stopped shipping in public. We partnered with Chainlink, NRN Agents, peaq. We integrated with Homebrew Robots, HuggingFace, and Coinbase's x402 protocol. World Labs, Rerun, MuJoCo creator and others endorsed what we were building. Not because we asked. Because they saw infrastructure that actually worked. We also did something no other Pumpfun project had done: we threw a real event at Solana Breakpoint. Robo Eve. Over 100 people showed up, met the team, and saw the vision in person. To our early supporters who held through every phase: this only exists because of you. To our advisors who kept us focused on what mattered, we’re grateful. To our team who just kept building, this wouldn’t be here without you. And to @Pumpfun: you gave builders with no connections a fair shot. Just a launch button and a community that decides what lives. We are not what we started as. We evolved to something better. We're scaling real deployments now, expanding SimArena, and building the execution layer that makes physical AI deployable. We’re building the infrastructure this space has been missing. Year two won't be easier. But it will be bigger. Happy anniversary, CodecFlow.

Four minutes. That's how long it takes to go from "I have a CAD file" to "my robot is running in a simulation" on SimArena. Most teams measure that journey in weeks. SimArena measures it in minutes.

Bessemer nailed it. 745 software companies raised $30M+ in the last 5 years. Robotics: 42. In a market 30x larger. That’s not a bubble, that’s decades of underinvestment correcting in real time. At the same time, “the API moment for robotics is a 2028+ story.” Right now, the companies actually making money are full-stack. Not just models, but data pipelines, deployment infra, hardware integration. The whole stack. What this really means is the window is open now, but only for teams building end-to-end. The verticalizers win first. Waiting for proof is how you miss both waves. That’s exactly why we’re building CodecFlow as one unified stack, not a bundle of disconnected tools.

CodecFlow's SimArena is going to be one of the defining platforms in robotics.

LLMs are just one way of representing the world. But the physical world isn’t made of language. It’s made of physics, interaction, and consequence. That’s why LLMs alone aren’t enough. We’re seeing a shift toward world models and VLA systems, where understanding and acting on the real world becomes the focus. That shift needs a different kind of infrastructure.

Latest SONIC whole-body control model from @NVIDIAAI, now in SimArena by CodecFlow. This enables: • physics-aware, whole-body execution • stable tracking of complex motion We’ve also added a policy manager to handle multiple policies in the same simulation. Bridging learned motion and real-world constraints.

Robots are starting to reason about the world, not just respond to it. Gemini ER introduces a layer that understands a goal, breaks it down, and plans before acting. That alone changes how these systems are built. Until now, most robotics stacks have been tightly coupled, with perception, planning, and control all tangled together. Hard to generalize, harder to scale. What we’re seeing now is a separation between a model that decides what should happen and a system that ensures it actually does. That’s the direction we’re taking at CodecFlow. Because knowing what to do is only half the problem. In the real world, timing drifts, sensors disagree, and actions don’t land the way you expect. So the stack matters. SimArena for environments, optr for continuous execution, and Fabric for low-latency inference close to where decisions happen. The intelligence is improving fast. Now it’s about how well it runs.

using mj_ray and mj_multiRay for better LiDAR performance and ability to support multiple types 2D, 3D, spinning, flash etc...

Robots don’t just see, they sense. Support for LiDAR, depth, force, and custom sensors brings us closer to real-world environments inside simulation. One step closer to SimArena v1.

With how quickly robotics and AI are advancing, robots becoming as common as phones and TVs is inevitable. When that happens, training them for your own needs shouldn’t be hard. That’s the direction CodecFlow is building toward.