Lambda Brings a Trillion-Parameter Model to MLPerf’s New Agent Test

Lambda has submitted Kimi K2.6, a model with more than one trillion parameters, to MLPerf Inference v6.1’s new agentic workload on an eight-GPU NVIDIA HGX B200 system. Lambda calls it MLPerf’s first deployment of a trillion-parameter model and the benchmark suite’s only agentic submission on datacenter hardware so far; the results are still awaiting final MLCommons review and publication.

The timely development is not simply a faster score. MLPerf v6.1 is the first round to include agentic tasks, which are built around multi-turn work where an AI system calls tools and each step depends on the earlier conversation. MLCommons says those sessions create distinct serving pressures: context grows over time, reusable key-value cache data becomes important, outputs vary in length, and one delayed turn can slow the entire task.

A benchmark built for dependent turns

That design shifts the measurement away from a single prompt-and-response exchange. The benchmark replays trajectories, or ordered sequences of requests, so a system must keep an agent moving through a task while preserving output-quality checks. MLCommons selected Kimi K2.6 and Qwen3.6-35B-A3B for separate evaluations rather than folding their outcomes into one combined score.

Why this is not an edge leaderboard

The result should not be read as a clean win over the closed-division edge submissions beside it. Lambda entered the open division because it replaced the reference workload’s smaller Qwen backbone with Kimi K2.6. The company used 1.44 TB of aggregate high-bandwidth memory across its B200 system to host a model it describes as roughly 40 times larger than the reference model.

That asymmetry is useful, but it answers a narrower question: what happens when the same general agent harness is given datacenter-scale memory and a frontier-scale model? It does not establish that the B200 configuration is inherently faster than edge machines, because the submissions differ in division, hardware and served model. The benchmark’s separate model methodology reinforces that distinction.

Lambda’s other v6.1 submissions

• GPT-OSS 120B on four NVIDIA Blackwell Ultra GPUs: Lambda reported 65,511 tokens per second offline and 58,195 tokens per second in the server scenario.
• The company said those GPT-OSS results improved 8.79% offline and 8.85% in the server scenario from its v6.0 results on identical hardware and benchmark conditions.
• Qwen3-VL-235B-A22B-Instruct: Lambda submitted its first MLPerf vision-language-model results on both HGX B200 and four-Blackwell-Ultra-GPU systems.

Three results, three different questions

The GPT-OSS figures are Lambda’s strongest direct round-over-round comparison because the company says the hardware and benchmark conditions were unchanged. It also says its offline score led the four-Blackwell-Ultra-GPU field by 2.38% over second place. Those claims remain company-reported while MLCommons completes its review.

The vision-language results widen the submission beyond text generation. Lambda reported 101.56 offline queries per second and 69.41 server queries per second for Qwen3-VL on HGX B200, and said those were respectively 28.5% and 2.29% above the fastest comparable submissions from the previous round. The comparisons are more specific than a general claim about multimodal inference: they concern this model, these scenarios and the cited prior-round results.

For infrastructure teams, the submission makes the changing shape of inference easier to see. Agent systems are not only judged by raw token output; they must handle growing conversational state and tool-dependent turns. Lambda’s open submission demonstrates that a large datacenter system can be put through that new workload. Its value as a comparison will depend on keeping the configuration boundaries visible when final MLPerf results are published.