Lambda Feature Trails

Feature trailActive

Agent Security Testing and Prompt Injection Defense

The platform now supports agentic experimentation loops, enabling AI agents to autonomously hypothesize, experiment, evaluate, and iterate on tasks. This includes structured experiment tracking via an API, leaderboards for performance comparison, idea branching for exploring different hypotheses, and history logging for maintaining context. The system is designed to leverage otherwise idle compute resources for continuous research and development. This post introduces AgentFlow, a novel framework for agent security testing and prompt injection defense, and discusses the transition from passive world modeling to active agents in physical AI.

5 posts · 2026

Timeline

Feature trailActive

Open Source AI Model Harness

The platform now supports NCode, an open-source AI model harness that provides transparency and control over how models operate, what tools they can access, and how their performance is measured. This allows for greater reliability and customization by enabling users to integrate their own models, fine-tune them for specific tasks, and monitor their execution. NCode addresses the 'black-box' problem of proprietary harnesses by offering a fully inspectable and modifiable infrastructure layer, facilitating the retargeting of agentic systems to local models. This post introduces Open Jarvis, a framework that decouples the agent harness from the specific LLM, enabling local models to achieve performance parity with cloud models at significantly lower cost and latency. It details a five-primitive spec (Intelligence, Engine, Agent Logic, Tools & Memory, Learning) that can be optimized for any model and hardware, using a frontier model for guided spec search and improvement. The post also provides concrete performance metrics and cost comparisons for running local models on Lambda's infrastructure.

2 posts · 2026

Timeline

Feature trailActive

Single-Image 3D Scene Reconstruction

The platform now supports PixARMesh, a novel method for single-image 3D scene reconstruction. This approach tokenizes image, object pose, and mesh into a single sequence, allowing an autoregressive Transformer to predict object poses and shapes token by token in one forward pass. It directly generates artist-ready meshes without relying on implicit representations or multi-stage pipelines, improving reconstruction quality, especially under occlusion. PixARMesh sets a new state of the art on the 3D benchmarks. This post introduces CVP (Central-Peripheral Vision-Inspired Multimodal Model), a novel approach for spatial reasoning in AI. CVP combines a target-affinity token for focusing on task-relevant objects with an allocentric grid for global spatial awareness, inspired by human vision. It achieves state-of-the-art results on 3D vision-language benchmarks like ScanQA, SQA3D, ScanRefer, Multi3DRefer, and Scan2Cap, demonstrating improved performance in 3D question answering, visual grounding, and dense captioning. The post highlights the importance of representation in spatial reasoning and showcases Lambda's GPU infrastructure for training such compute-intensive multimodal workloads.

3 posts · 2026

Timeline

Feature trailActive

MLPerf Training Benchmark Performance

This capability tracks Lambda's performance in MLPerf benchmarks, showcasing advancements in LLM and MoE training and inference speeds on cutting-edge NVIDIA hardware. It highlights software optimizations and system design impacts on training and inference efficiency, providing a baseline for real-world AI training and inference workloads. The v6.1 submission demonstrates significant speed improvements for GPT-OSS 120B on 4 NVIDIA Blackwell Ultra GPUs and introduces the first agentic inference workflow. The EGInterpolator research, accepted at ICLR 2026, demonstrates a novel approach to AI-driven molecular dynamics for drug discovery by first learning molecular structures from abundant static data before fine-tuning on scarce dynamic trajectory data. This method significantly improves the accuracy of molecular motion prediction and reduces reliance on expensive simulations, showcasing the power of GPU-native AI workloads for scientific computing.

3 posts · 2026

Timeline

Feature trailActive

Agentic Experimentation Platform

The platform now supports agentic experimentation loops, enabling AI agents to autonomously hypothesize, experiment, evaluate, and iterate on tasks. This includes structured experiment tracking via an API, leaderboards for performance comparison, idea branching for exploring different hypotheses, and history logging for maintaining context. The system is designed to leverage otherwise idle compute resources for continuous research and development. This post introduces the_lab.api, an API-first experiment tracking system that allows agents to manage experiments, compare results, and branch ideas programmatically, enabling autonomous research cycles. It also highlights the use of spare GPU capacity for these experiments, demonstrating cost-effective utilization.

3 posts · 2026

Timeline

Feature trailActive

3D Object-Centric Scene Representation Learning

The platform now supports 3D-DLP, a novel self-supervised approach to decompose colored 3D scenes into object-centric particles. This method extends Deep Latent Particles (DLPs) to 3D, enabling the prediction of object position, size, and appearance without relying on annotations or pre-trained segmentation models. It addresses challenges in 3D scene modeling, such as the visual binding problem caused by tokenization, by learning representations of objects directly in the latent space. The framework now also incorporates StereoPolicy, which enables robot policies to learn 3D spatial understanding directly from stereo images by fusing features from pretrained 2D vision encoders through a cross-attention Stereo Transformer, capturing 3D structure implicitly without explicit depth map or point cloud estimation. This approach has shown improved success rates on real-world tabletop tasks, particularly with transparent objects and in low-data regimes.

3 posts · 2026

Timeline

Feature trailActive

AI Model Deployment Strategy and Cost Optimization

This capability focuses on developing strategies for AI model deployment, emphasizing a hybrid approach that balances the use of frontier models with open-weight models. It addresses the economic challenges of AI spend by advocating for cost-effective allocation of workloads, leveraging reserved compute for predictable costs, and enabling fine-tuning of open-weight models on proprietary data for competitive advantage. The system aims to provide organizations with a framework for optimizing AI in production by enabling them to choose between proprietary APIs and self-hosted open-weight models based on workload requirements and cost-effectiveness. This post introduces the concept of a 90/10 split, recommending open-weight models for the majority of tasks and frontier models for specialized, high-demand applications. It also highlights the benefits of owning the model, including cost predictability, control over the roadmap, and enhanced privacy through fine-tuning on proprietary data.

1 post · 2026

Timeline

Feature trailActive

Adaptive Data Preprocessing for Financial Markets

This capability focuses on developing adaptive data preprocessing techniques for high-frequency trading (HFT) environments. It addresses the challenge of non-stationary market conditions by moving beyond fixed normalization strategies. The core idea is to build systems that learn market structure directly from data, enabling them to preprocess and adapt to incoming data dynamically, even across unseen market conditions. This involves using Bayesian frameworks to model market behavior and statistical divergence to distinguish between noisy and novel data, thereby reducing the compute burden of the training loop and improving the efficiency of GPU infrastructure.

2 posts · 2026

Timeline

Feature trailActive

Kubernetes AI Workload Scheduling

The platform now supports multiple orchestration layers for AI workloads, including Slurm for batch training, Kubernetes for flexible containerized workloads, SkyPilot for portable job execution, and dstack for vendor-agnostic management. This allows teams to align the platform with their specific requirements for scheduling, resource allocation, and configuration overhead. This post details the limitations of the default Kubernetes scheduler for AI workloads, specifically its lack of gang scheduling and multi-node fabric topology awareness, and introduces Kueue, KAI Scheduler, and Volcano as solutions, along with common deployment patterns like Kueue + Volcano, KAI Scheduler Standalone, and Volcano Standalone.

2 posts · 2026

Timeline

Feature trailActive

Dynamic GPU Resource Scaling for ML Competitions

The platform now supports dynamic scaling of GPU resources, allowing additional GPUs to be added mid-competition without interrupting ongoing battles. This is achieved by independently serving models on each GPU with vLLM, using Caddy for load balancing and health checks, and a dynamic concurrency limiter that resizes based on available resources. This enables efficient cost management and ensures competition integrity even with fluctuating demand. This post highlights the increasing demand for serving large open-weight models like GLM 5.2, emphasizing the infrastructure challenges and the need for efficient serving solutions to keep up with the capability gap closing between open and closed-weight models.

2 posts · 2026

Timeline

Feature trailActive

Large-Scale Synthetic Data Generation

The platform now supports large-scale synthetic data generation, enabling the creation of diverse and high-signal training data for AI models. This includes systems for procedural generation of physical scenarios using simulators, automatic construction of verified question-answer pairs, and optimization of data generation pipelines for various model sizes and benchmarks. The focus is on moving beyond human annotation to scalable, simulation-driven data creation. The Sim2Reason system, accepted at ICML 2026, demonstrates that LLMs can learn to solve complex physics problems using only synthetic data generated from physics simulators. This system procedurally generates diverse physical scenarios, extracts verified question-answer pairs (numeric, reverse, and symbolic), and has shown significant zero-shot performance improvements on physics and math reasoning benchmarks. This capability is crucial for addressing the limitations of real-world data and the increasing demand for synthetic data in AI training.

1 post · 2026

Timeline