September 2026 Summaries

Long-horizon agentic reinforcement learning is often limited not by raw model token speed but by the time required to assemble a fresh, valid training batch from trajectories that repeatedly alternate between generation, tool execution, sandbox operations, and verification. Experiments using GLM-5.2 on 32 GPUs showed that completion-order scheduling, surplus candidate groups, cache-aware sticky routing, isolated sandbox execution, optimized KV-cache storage, and speculative decoding reduced fixed-weight Terminal-Bench batch collection for 256 trajectories from 591.8 to 196.1 seconds, a 66.9% improvement. The system also avoided global stalls during policy updates by rotating updates across serving engines and preserving eligible partial work under explicit freshness limits, producing a complete Terminal-Bench interval of 240.1 seconds and limiting the added cost of live weight updates on SWE-Bench to roughly 11%. The findings argue that agentic RL infrastructure should measure time to a trainer-consumable batch rather than inference throughput alone, while balancing concurrency, cache locality, sandbox contention, and controlled policy staleness.

Nebius and WEKA benchmarked WEKA NeuralMesh with Augmented Memory Grid, a distributed KV-cache tier that pools NVMe storage across an eight-node NVIDIA HGX B300 cluster, to reduce repeated LLM context processing in long-running agentic coding sessions. In an eight-hour replay of real traffic against DeepSeek-V4-Pro, using identical hardware, vLLM configuration, active-session pool, and workload for both tests, the shared cache configuration served 55,694 requests versus 23,015 for GPU HBM alone, achieved a 93% cache-hit rate versus 40%, and reduced errors from 2.32% to 0.64%. It also lowered median time to first token from 5.68 seconds to 1.05 seconds and p95 latency from 23.12 seconds to 3.82 seconds, while producing 3.7 times more output tokens per GPU-hour. The results indicate that the shared tier maintained stable performance as contexts grew toward 900,000 tokens, whereas the HBM-only setup experienced rising latency and timeout rates as its working set exceeded GPU memory capacity.

SemiAnalysis awarded Nebius its highest ClusterMAX 3.0 Platinum rating after months of testing live NVIDIA GB300 and HGX B300 clusters under training, inference, storage-concurrency, and recovery workloads. The assessment highlighted Nebius’s reliable operations, topology-aware Slurm scheduling, current software packages, network performance, and low-friction orchestration, while recognizing its open-source Soperator platform for combining Slurm scheduling with Kubernetes and simplifying storage management through VirtioFS-backed persistent volumes. Testing identified storage performance and metadata-consistency issues under heavy concurrent use, which Nebius addressed before retesting confirmed resolved errors, above-median per-client performance, and scalability across two racks. SemiAnalysis also found that Nebius automatically detected, rebooted, and restored a failed node in a GB300 NVL72 rack without manual intervention, and praised its monitoring dashboards, automated remediation, and cluster-health visibility.

Nebius AI Cloud is introducing dynamic spot pricing for preemptible VMs, which use temporarily idle GPU capacity for interruptible, non-critical workloads such as data processing, experiments, and short fine-tuning jobs. The system automatically adjusts prices based on available capacity and customer demand, with prices updated as often as every 15 minutes and constrained between a floor and one cent below the on-demand GPU-hour rate. Customers can review up to 30 days of pricing history and either set a maximum acceptable price, causing workloads to stop and retain data when prices exceed that threshold, or follow the current spot price to prioritize continued access to capacity until maintenance or capacity reclamation occurs. The model aims to reduce idle data-center compute, improve energy efficiency, and give users lower-cost GPU access, while dynamic pricing will take effect on October 8 and existing preemptible-resource users must configure a pricing setting before then.

Nebius reported its broadest MLPerf Inference v6.1 submission, benchmarking seven NVIDIA-based system configurations across DeepSeek R1, Qwen3-VL 235B, and gpt-oss 120B in server and offline scenarios. The company earned five first-place results among 20 category entries, led by its 72-GPU GB300 NVL72 rack configuration, which achieved first-place DeepSeek R1 throughput of 603,023 tokens per second in the server scenario and 689,961 in offline, while exceeding 1.1 million tokens per second on gpt-oss 120B. Nebius also submitted preview results for a 36-GPU Vera Rubin NVL72 system, one of only two Vera Rubin-based submissions, reporting 16,427 server tokens per second per GPU for DeepSeek R1. Results indicated near-linear scaling from eight to 72 GB300 GPUs for gpt-oss 120B, performance gains from HGX B200 to B300 systems, and competitive results from an eight-GPU RTX PRO 6000 configuration positioned for cost-efficient inference.

Nebius has launched the free AI Builder Program to support developers building AI systems with open models and interoperable tools rather than relying on a single vendor. The program provides more than $400 in credits and discounts from Nebius Token Factory, LangChain, Toloka, Tavily, and other partners, along with working-code cookbooks, training through Nebius Academy and NVIDIA, discounted certifications, expert office hours, and a builder community on Discord and at live events. Its approach centers on “open primitives,” enabling users to combine models, inference, orchestration, retrieval, evaluation, coding agents, human labeling, and post-training tools into customizable systems that can be updated as new models emerge. Initial partners include NVIDIA Nemotron, Qwen, MiniMax, Cognition, OpenHands, LangChain, Tavily, Composio, Toloka, LangSmith, Prime Intellect, and Hugging Face OpenEnv, while Nebius supplies infrastructure for hosting, serving, training, and fine-tuning open models.

Nebius submitted MLPerf Storage v3.0 Closed division results for its Enhanced Object Storage service, using standard CPU-only virtual machines and the open-source mlpstorage harness to simulate AI-training I/O over the S3 API. The company reported feeding 768 simulated NVIDIA B200 accelerators for RetinaNet, the highest accelerator count in the round, with 88.4% utilization, 102.83 GiB/s read bandwidth, and 341,881 samples per second; it also achieved the highest Unet3D accelerator count among object-storage submissions, sustaining 114.45 GiB/s and 92.5% utilization across 21 simulated accelerators. For Llama 3.1 checkpointing, aggregate write bandwidth rose from 2.81 GiB/s on one node for an 8B model to 19.46 GiB/s across eight nodes for a 70B model, while read bandwidth increased from 7.14 GiB/s to 38.27 GiB/s. MLPerf Storage evaluates whether storage can maintain minimum accelerator utilization under realistic training workloads, with RetinaNet emphasizing small-file latency and Unet3D emphasizing large-file bandwidth, and Nebius presents its results as evidence that a single production object-storage service can support training and checkpointing workloads at scale.