CoreWeave claims fastest Kimi K2.6 inference among 11 providers

Fastest output speed in the benchmark

According to CoreWeave, its inference service delivers the highest token throughput and lowest latency among providers serving Kimi K2.6. The company points to a chart of output tokens per second for a 10,000-token input, accurate as of 5/11/2026. CoreWeave tops the chart at 205 tokens per second. The next values shown are 158 and 125, and the lowest is 22. The chart lists providers including Clarifai, Azure, Cloudflare, Fireworks, SiliconFlow (FP8), Novita, Together.ai (FP4), DeepInfra (FP4) and Parasail.

Artificial Analysis evaluates output speed in tokens per second and price per million tokens, which it calculates as a 7:2:1 blend of cache-hit, input and output costs. Its Speed vs. Price view for Kimi K2.6 compares 11 providers. CoreWeave describes this as an independent third-party measurement and says it is the leading provider in the high-speed, low-price quadrant today.

About Kimi K2.6

Kimi K2.6 is Moonshot AI's flagship open-source model, released on April 20, 2026. CoreWeave says it is one of the most widely used open-weight models, ranking #2 on OpenRouter's leaderboards. It is a 1-trillion-parameter Mixture-of-Experts model with 32 billion active parameters per token, native multimodal input and a 262K context window.

CoreWeave describes the model as competitive with leading proprietary frontier models, citing scores of 86.3 on BrowseComp, 80.2 on SWE-Bench Verified and 66.7 on Terminal-Bench 2.0. The company adds that it has become a favored model for coding agents and multi-step agent frameworks.

NVFP4 and EAGLE3 on NVIDIA GB300

CoreWeave credits its Applied Training team, an internal research group, for the result. The team uses the company's NVIDIA GB300 and GB200 NVL72 clusters and techniques such as NVIDIA Model Optimizer PTQ. According to CoreWeave, training a custom NVFP4 quantized model together with EAGLE3 speculative decoding produced the greatest performance gains on Kimi K2.6.

The company says it validated model quality on Terminal-Bench 2.0, SWE-Bench Verified, SWE-Bench Pro, GPQA Diamond, AIME 2026 and LiveCodeBench v6, using Weights & Biases and test harness frameworks. In its words, validating the NVFP4 weights is essential for improving performance without degrading quality. It also says it tunes the whole stack, from bare-metal access to NVIDIA accelerators and high-speed interconnects to memory architectures and custom inference stack tuning.

Deployment options and what it means for customers

CoreWeave argues that in production inference every performance lever also affects cost, including GPU selection, quantization, attention backend, KV cache layout, batch sizing, speculative decoding and kernels. Real-time customer-facing workloads need low latency, while offline jobs such as evaluations need high throughput at low cost.

• Serverless Inference: the fastest path to deployment, with many models behind a single unified API.
• Dedicated Inference: explicit GPU selection, runtime configuration and auto-scaling, with CoreWeave managing the infrastructure.
• Inference on CoreWeave Kubernetes Service (CKS): direct control over GPUs, runtimes, orchestration and capacity management.

For teams building coding agents or other agent workloads on Kimi K2.6, the claim is that they can get higher output speed at a competitive price, and choose between token-based consumption and GPU-based scaling. The ranking is a snapshot as of the date on the chart, and the source does not give prices.