Lambda

Our co-founder and CTO Stephen Balaban joins MTS Live in ~20 minutes to talk open weights, fine-tuning, and what comes next. Tune in: https://lnkd.in/gDNvFSTk

Open models move science forward. When AI engineers can download, inspect, and tune a model for their needs, progress stops being gated by who owns the weights. Research and open source have been core to Lambda from the start. That's why we have signed the Open Weights letter alongside NVIDIA, Microsoft, and 230+ others.

Run 100,000 battles of untrusted, adversarial agent code. Keep every one of them in its own lane. That's the infrastructure problem behind AgentBeats, a competition where attacker agents and defender agents try to break each other. The hard part isn't scoring the match. It's running code you don't trust, at scale, without one battle leaking into the next. Here's how Lambda kept 100k of them isolated, and what happened when the agents went at each other: https://lnkd.in/gbgg6fbd

In high-frequency trading (HFT) data, noise isn't the problem. Assumptions are. Every quant team hits the same trap: a preprocessing pipeline that's either too rigid (overfit to old regimes) or too manual (constant engineer babysitting). Neither holds at HFT scale. This is a representation-learning problem disguised as an engineering challenge. The question isn’t about cleaning data. It’s about building a system that understands the data well enough to preprocess it automatically. Part one of a two-part series. Read more: https://lnkd.in/gvNy6iq3

Your coding tools shouldn't be a black box. The model is trained together with the software that drives it. Noumena ran their whole lab on NCode for six months: about 25M lines of code, dozens of services, semi-supervised agents. It's open source. Bring any model and point it at an OpenAI-Compatible endpoint. You can't cleanly split the model from the software driving it. A smaller model with the right tooling will often beat a larger one running blind. NCode gives you a tuned setup you can actually read and change. Every model earned its tier through months of real use across live workloads. Learn more about why NCode was developed, tested, and the efforts Noumena went through to productionize open-weight models in our blog: https://lnkd.in/gp36htiY

Lambda is among the first adopters of NVIDIA Spectrum-X with co-packaged photonics, the industry's first CPO switch in volume manufacturing. Vera Rubin is here, and it's ramping up to gigascale. Read more: https://lnkd.in/dx5d729P

Point a camera at a room. Get back a full 3D model of everything in it: the sofa, the table, the chairs. Each one a clean mesh you can rotate, move, and drop straight into a game engine. A single photo shows one viewpoint. Half the scene is hidden, depth is ambiguous, objects overlap. Reconstructing accurate 3D from that has been an underdetermined problem for years. PixARMesh (a CVPR 2026 method from UC San Diego and Lambda) borrows a trick from language models. Tokenize the image, the object poses, and the meshes into one sequence, then let an autoregressive Transformer predict it token by token. One forward pass, no signed distance fields, no separate layout-optimization loop. On 3D-FRONT, it sets a new scene-level state of the art (F-Score 33.55% vs 25.00% for the prior best), and it outputs compact, artist-ready meshes instead of dense, over-tessellated surfaces. Read more and explore the interactive demo: https://lnkd.in/gyBjBr-w

Your Kubernetes scheduler is quietly wasting GPUs. A distributed training job asks for 16 GPUs. Four are free. The default scheduler grabs those four, then sits and waits for the other twelve. Nobody else can use them. Your job isn't training. The cluster is busy doing absolutely nothing. It's a partial-scheduling deadlock. At scale, it breaks the whole workflow. We compared three schedulers built for AI workloads, Kueue, KAI Scheduler, and Volcano, on the two things that actually matter: gang scheduling and topology awareness. See which one fits your cluster: https://lnkd.in/gnaHFG7a

Open-weight models just hit a tipping point. GLM 5.2 runs as a real subagent in production. Not as a demo. Not as a benchmark screenshot. But as a near-frontier performance you can actually deploy. Intelligence is becoming a commodity. The bottleneck moves to infrastructure, to who can serve it, at what throughput, and at what cost. Running GLM 5.2 unquantized on-premises costs hundreds of thousands of dollars, and the throughput still won't match that of a real cluster. We broke down where GLM 5.2 lands and why the hard part isn't the model anymore. Read more: https://lnkd.in/gJj22PSG

Lambda has been recognized in the first Gartner® Magic Quadrant™ for Cloud AI Infrastructure, based on completeness of vision and ability to execute. It's also the first time Lambda has been recognized in one. This recognition reflects our focus: delivering against our commitments. Read the report: https://lnkd.in/gBYbgA65