September 2026 Summaries

Open Jarvis is an open-source framework designed to improve locally run open-weight AI models by optimizing the surrounding agent harness—such as prompts, inference runtime, tool access, memory, and reasoning logic—for a specific model and hardware setup. Its research reports that a Qwen3.5-9B model’s PinchBench accuracy rose from 62.3% to 88.4% after harness retargeting, recovering much of the gap to Claude Opus 4.6 without changing the underlying model. The system separates configuration into Intelligence, Engine, Agent Logic, Tools and Memory, and Learning, allowing automated spec search guided by a frontier cloud model to identify changes that improve performance while avoiding regressions. Open Jarvis argues that this modular approach enables local models to approach cloud-model performance on many personal AI tasks with lower query costs and latency, while retaining control over data and infrastructure. On Lambda’s NVIDIA HGX B200 hardware, an evaluation found that self-hosted Kimi K2.6 scored below Claude Fable 5.1 on shared PinchBench tasks but had an estimated 34% lower per-task inference cost, illustrating the potential tradeoff between quality, cost, and operational control.

Lambda announced plans to build a data center in MidAmerica Industrial Park in Mayes County, Oklahoma, projecting up to 1,000 construction jobs, up to 100 permanent jobs, and an estimated $500 million in tax revenue over 10 years under current rates. The company says it will pay all of the facility’s electricity and required infrastructure costs in accordance with Oklahoma ratepayer-protection law, pursue flexible computing loads during peak demand, and use closed-loop cooling to reduce water consumption. Located in the industrial park’s southern section within the Chouteau-Mazie Public School District boundary, the project is expected to support local schools, public services, infrastructure, vendors, and contractors. MidAmerica, GRDA, and Cloverleaf Infrastructure selected and are preparing the site, while Lambda and MidAmerica plan to continue consulting local leaders and residents as development planning and project timelines progress.

Spatial reasoning enables AI systems to understand 3D environments, including object locations, relationships, and task-relevant context, an ability needed for robotics and other physical AI applications. Researchers at UC San Diego and Lambda developed CVP, a multimodal model inspired by human central and peripheral vision that combines a target-affinity token for identifying relevant objects with an allocentric, bird’s-eye-view grid for representing the wider scene. Built on LLaVA-Video-7B and Video-3D-LLM, CVP integrates multi-view visual features, 3D positional information, and language, and was trained across five datasets covering question answering, visual grounding, and dense captioning. On benchmarks including ScanQA, SQA3D, ScanRefer, Multi3DRefer, and Scan2Cap, it outperformed the Video-3D-LLM baseline, with its object-focused mechanism especially aiding grounding tasks and its global grid particularly supporting scene-based questions and descriptions. The work suggests that effective spatial AI depends not only on richer geometric data but also on representations that balance local attention with global environmental awareness, while requiring substantial multimodal training infrastructure.

Molecular dynamics simulations help drug discovery researchers evaluate how drug candidates and protein targets behave over time, but their high computational cost limits their use at scale. EGInterpolator, developed with Stanford University and accepted at ICLR 2026, addresses the shortage of expensive molecular dynamics trajectory data by first pretraining a diffusion model on abundant static three-dimensional molecular conformer data, then fine-tuning it on dynamics data to generate trajectories. On the DRUGS benchmark, the approach outperformed the prior GeoTDM method, reducing divergence from reference simulations by approximately 73% for bond angles, 78% for bond lengths, and 24% for torsional motion, while ablation tests indicated that structural pretraining was central to these gains. The method was also extended to tetrapeptides and protein monomers, suggesting a data-efficient path for AI-driven molecular simulation in drug discovery, supported by GPU-based infrastructure such as Lambda’s NVIDIA-powered training environment.

EdiVal-Agent is an agentic AI framework developed by Lambda with researchers from the University of Texas at Austin, UCLA, and Microsoft to automate evaluation of instruction-based image editing models, addressing the difficulty of assessing whether edits follow requests, preserve unrelated content, and maintain visual quality across multiple turns. Accepted at ICLR 2026, it decomposes editing instructions into object-level requirements, maintains an evolving record of objects and prior edits, and combines vision-language reasoning, object detection, verification rules, and human-preference models to assess instruction following, content consistency, and visual quality. Its instruction-following component achieved 81.3% agreement with human judgments, exceeding VLM-only and CLIP-based evaluators, while benchmarks showed that models performing well on single edits can still degrade as sequential instructions accumulate. The framework is intended to support continuous model development by identifying specific tradeoffs and failure modes across checkpoints, illustrating how AI agents can serve not only as user-facing tools but also as scalable infrastructure for evaluating and improving increasingly complex multimodal foundation models.

Lambda reports MLPerf Inference v6.1 results highlighting advances in large-language-model, vision-language-model, and agentic inference on NVIDIA Blackwell-based systems. Its four-NVIDIA Blackwell Ultra GPU configuration achieved 65,511 tokens per second offline and 58,195 tokens per second in the server scenario for GPT-OSS 120B, leading comparable offline submissions and improving throughput by roughly 8.8% over its v6.0 results on identical hardware through software optimizations. Lambda also submitted Qwen3-VL-235B-A22B-Instruct, its first vision-language-model benchmark, with an HGX B200 system reaching 101.56 offline queries per second and 69.41 server queries per second. In the open division, it used an eight-GPU HGX B200 system to run Kimi K2.6, a mixture-of-experts model with more than one trillion parameters, in MLPerf’s new agentic workload, completing all 1,007 replay turns with 86.83% BFCL v4 accuracy and a mean per-turn latency of 770.8 milliseconds. The company attributes performance gains to both Blackwell Ultra hardware capabilities and optimizations including TensorRT enhancements, CUDA graphs, model quantization, KV-cache tuning, and automated expert-kernel selection, while noting that all results remain subject to final MLCommons review.

Lambda announced on September 14, 2026, that it signed the White House Ratepayer Protection Pledge, committing to develop AI infrastructure in ways intended to benefit local communities while protecting electricity consumers. The company said it will incorporate community input early in project planning, support policies requiring data center operators to pay for their power use, and advocate for expanded power generation, grid improvements, and reliable service as AI-related electricity demand grows. Lambda also stated that it invests in technologies intended to reduce local resource demands. Founded in 2012, Lambda provides AI cloud infrastructure and supercomputers for training and inference to researchers, enterprises, and hyperscalers, while noting that its statements about future plans and performance are subject to risks and uncertainties.

OpenResearcher, an EMNLP 2026 project by researchers from several universities and Lambda, addresses the challenge of training deep research agents by generating reproducible synthetic research trajectories rather than relying on expensive human demonstrations or unstable live-web data. It creates an offline environment containing 15 million documents and search, open, and find tools, where GPT-OSS-120B generated more than 97,000 long-horizon trajectories involving planning, retrieval, evidence gathering, failed tool calls, and iterative reasoning. These trajectories were filtered and used to fine-tune the smaller NVIDIA Nemotron-3-Nano-30B-A3B model with long-context training, producing an agent that achieved 54.8% on BrowseComp-Plus and substantially outperformed several larger proprietary and open-source systems on reported benchmarks. Although trained entirely offline, the model also generalized to live-web-oriented evaluations, supporting the claim that large-scale synthetic research experience can transfer to real-world research tasks. The project illustrates a shift toward allocating substantial compute to synthetic-data generation, with teacher inference using 64 H100 GPUs for two days compared with eight H100 GPUs for eight hours of student post-training, and its trajectories have since been adopted in NVIDIA research and synthetic-data tooling.

StereoPolicy is a robotics perception approach developed by researchers at Stanford, Northwestern, and Lambda that learns implicit 3D spatial understanding from synchronized stereo image pairs rather than relying on monocular vision, depth maps, or point clouds. Using pretrained 2D vision encoders and a cross-attention Stereo Transformer, it identifies correspondences between left and right images to support manipulation tasks involving clutter, thin structures, and transparent objects without constructing an explicit geometric representation. The system can be integrated into diffusion policies trained per task or fine-tuned vision-language-action models, and tests across real-world tabletop tasks, simulations, and low-data VLA settings showed higher success rates than RGB, RGB-D, and point-cloud baselines, including on the difficult task of hanging a glass cup. The work is positioned within the growing physical AI field, where scalable GPU infrastructure is increasingly important for training and deploying foundation-model-based robotic systems.

SPREEAI’s virtual try-on system generates photorealistic images of shoppers wearing catalog garments by jointly addressing pose transfer, cloth deformation, identity preservation, and texture and logo fidelity, rather than relying on simple two-dimensional garment warping. After initially using separate models that caused cascading errors, the company consolidated its pipeline into a unified diffusion model trained on high-resolution imagery and validated through identity, texture, and cloth-physics checkpoints, requiring the 80 GB memory capacity of NVIDIA H100 GPUs. Its infrastructure work included high-bandwidth distributed training, stage-aware data sharding, S3-based artifact retention, and MLflow-driven scheduling and observability, which reduced storage costs by more than 30 times, eliminated storage stalls, and raised GPU utilization from roughly 20% to 43%. For production inference, the team found that weight-only INT8 quantization saved memory but did not accelerate tensor-core computation, so it adopted true FP8 linear layers and a Hopper-optimized NVIDIA cuDNN attention kernel; together, these changes reduced denoising latency from up to 3.35 seconds to 1.97 seconds, more than halved peak memory use, and allowed approximately eight to eleven concurrent requests on one H100 GPU. The company plans to extend the system to video try-on, which will require temporal consistency, larger-scale compute and networking, and potentially tensor-parallel training, while also developing visual size prediction and broader commerce applications.

ReviewGrounder is an AI-assisted peer-review system designed to address growing reviewer workloads while keeping final judgment with human reviewers, area chairs, editors, and researchers. Rather than relying on a single model to assess an entire paper, it uses a Review Drafter followed by specialized agents that investigate related literature, technical reasoning, and experimental evidence, before an aggregator combines their findings with a conference rubric into a cited, evidence-based review. The system aims to make critiques more specific and verifiable by linking comments to relevant paper sections, prior work, and evaluation criteria. According to results reported for ACL 2026, ReviewGrounder exceeded GPT-4.1 by 41% and GPT-4o by 135% across eight review-quality dimensions. Developed through a collaboration involving several universities and Lambda, the project is presented as part of a broader movement toward compute-intensive AI research agents that support scientific writing, experimentation, and evaluation.