Artificial intelligence news.
Straight from the original announcements.

CoreWeave Brings Kimi K2.7 Code to Serverless Inference, Tops Artificial Analysis Chart

October 2, 2026, 16:12By the AI News Bites Editorial TeamSource: CoreWeave
CoreWeave Brings Kimi K2.7 Code to Serverless Inference, Tops Artificial Analysis Chart
Image: CoreWeave

CoreWeave has made Moonshot AI's Kimi K2.7 Code available on its Serverless Inference platform. According to the company, it delivers the highest output speed at a low blended price among providers serving the model, as measured by Artificial Analysis.

Key points
  • Kimi K2.7 Code from Moonshot AI is now on CoreWeave Serverless Inference.
  • According to CoreWeave, it leads Artificial Analysis's Speed vs. Price chart for the model, with the highest output speed at a low blended price.
  • CoreWeave converted the INT4 weights to NVFP4 and trained a DFlash speculative decoder on over 1 million prompts.
  • The D-PACE loss function gave an approximate 8% improvement in acceptance length, per CoreWeave.
  • The model is also available via Dedicated Inference and CoreWeave Kubernetes Service.

What CoreWeave is claiming

CoreWeave says it sits in the most attractive quadrant of Artificial Analysis's Speed vs. Price chart for Kimi K2.7 Code. According to the company, it delivers the highest token throughput with the lowest latency among all providers serving the model, which it describes as the best price-performance.

Artificial Analysis measures output speed in tokens per second and price per million tokens. For price, it uses a 7:2:1 blend of cache-hit, input and output costs. CoreWeave argues this reflects real workload distribution more accurately than input or output price alone.

CoreWeave says this is the second Kimi model in a row it has optimized and put live with improved performance. It earned the same validation on Artificial Analysis for Kimi K2.6, which it covered in an earlier blog post.

About the model

Kimi K2.7 Code is described as Moonshot AI's latest coding-focused agentic model. It is built on the same trillion-parameter Mixture-of-Experts architecture as Kimi K2.6, with 32 billion active parameters per token and a 256K-token context window. The open weights are released under a Modified MIT license. The model is tuned for long-horizon, end-to-end coding tasks and agentic tool use.

Moonshot AI reports that Kimi K2.7 Code uses about 30% fewer reasoning tokens than K2.6 on the same work. CoreWeave notes that this cuts cost and latency in agent loops that call the model repeatedly.

The optimization work: NVFP4 and DFlash

CoreWeave's Applied Training team used NVIDIA GB300 NVL72 and GB200 NVL72 clusters, along with NVIDIA's Model-Optimizer for post-training quantization. The model shipped pre-quantized in INT4. Because Blackwell accelerates FP4 formats such as NVFP4 rather than integer formats, CoreWeave says the weights needed to be requantized to benefit fully from Blackwell hardware.

  • Baseline: the team measured the INT4 weights on evals including Terminal-Bench 2.0, AIME2025, SWE-Bench Pro and SWE-Bench Multilingual, using the Harbor framework with the Terminus-2 agent.
  • Quantization: a custom 3-pass calibration converted the weights to NVFP4. It combined a short-context pass over general conversation data with long-context passes over math and coding datasets.
  • Speculator: the team trained a DFlash speculative decoding model on a dataset generated from the NVFP4 checkpoint, using over 1 million prompts spanning long-context agentic tasks, tool calling and coding work.

For training, CoreWeave used the TorchSpec project from the LightSeek Foundation with a modified loss function called D-PACE, which came from Harvard and MIT researchers. According to CoreWeave, it gave an approximate 8% improvement in acceptance length across all categories as measured by the SPEED-Bench AL dataset. The company says it has contributed D-PACE support back to TorchSpec upstream, and added support for commonly reported acceptance length evals to NVIDIA's AIPerf.

Draft model training ran on multiple GB300 NVL72 racks managed by the CoreWeave Kubernetes Service over a RoCE fabric. The production endpoint runs the NVFP4 plus DFlash configuration on vLLM by default, with no additional setup for customers.

What it means for customers

CoreWeave stresses that model-level gains are amplified by the rest of its stack, from bare-metal access to NVIDIA GPUs and high-speed interconnects to memory architectures and custom inference tuning. It says customers get the price-performance gains across all inference deployment paths.

  • Serverless Inference: one API, pay per token, with the optimizations included and no infrastructure to run.
  • Dedicated Inference: more control over GPU selection, runtime configuration and auto-scaling, with CoreWeave managing the infrastructure.
  • CoreWeave Kubernetes Service (CKS): direct control over GPUs, runtimes, orchestration and capacity.

The results are the company's own claims, backed by the Artificial Analysis chart it cites. Teams running coding agents on open-weight models can try Kimi K2.7 Code on CoreWeave Inference now.

Related news

Products and tools
Products and toolsNovember 13, 2026, 22:09

CoreWeave launches one-click GPT-J inference service with scale-to-zero

CoreWeave says it is making EleutherAI’s GPT-J available through a one-click inference service, combining the open-source model with access to its GPU infrastructure.…

AWS Makes SGLang-Neuron Publicly Available for Trainium
Chips and infrastructureNovember 12, 2026, 23:03

AWS Makes SGLang-Neuron Publicly Available for Trainium

AWS says SGLang-Neuron is now publicly available through an AWS Deep Learning Container released by RadixArk in collaboration with Annapurna Labs. The plugin brings…

Moonshot AI’s Kimi K3.1 Pro Highspeed arrives on Amazon Bedrock
ModelsNovember 11, 2026, 02:07

Moonshot AI’s Kimi K3.1 Pro Highspeed arrives on Amazon Bedrock

Moonshot AI’s Kimi K3.1 Pro Highspeed is now available on Amazon Bedrock, giving developers access to a post-trained Kimi K3 model with a 1-million-token context window…