.png)
Beijing and San Francisco — Moonshot AI, the company behind the Kimi AI assistant and Kimi family of models, and Inferact, the company founded by the creators of vLLM, today announced an expanded collaboration to support the Kimi model series in vLLM, including Kimi K3 and K3.1 Pro, spanning pre-launch model tuning, native vLLM integration, and deployment validation.
The teams behind Inferact and Kimi have worked together for years to bring Kimi’s open-weight model releases to the vLLM ecosystem. With K3, that collaboration expanded beyond open-source integration to include pre-launch tuning and production qualification.
K3.1 Pro, released today, is available on the Inferact Platform from launch. The teams held it to the same rigorous standard they applied to K3, which Inferact brought onto its platform after completing final qualification and Kimi Vendor Verification.
Together, the companies are connecting model development, open-source systems engineering, and production operations so improvements can move more quickly from an early checkpoint to dependable service.
Open-source collaboration through vLLM
Moonshot AI and Inferact are also collaborating to make the Kimi models first-class citizens in vLLM. Ahead of K3’s release, the teams coordinated on its architecture and serving requirements so developers could use a native open-source implementation from day zero rather than waiting for a separate integration after launch.
Supporting K3 required systems work tailored to its scale and architecture, including Kimi Delta Attention, hybrid cache management, native vision inputs, reasoning and tool-call parsers, structured output, prefill-decode disaggregation, and deployment across heterogeneous hardware. Inferact also developed and trained a speculative decoding model to increase generation throughput on production-class systems.
The work continued after launch as production measurements informed further improvements to scheduling, caching, memory use, and kernel performance. Moonshot AI and Inferact are bringing the same early coordination to K3.1 Pro so its open-source serving path can evolve alongside the model.
Production serving on Inferact Platform
Inferact serves Kimi K3 and K3.1 Pro in production on its platform and works with Moonshot AI to contribute the optimizations from that work back to the open-source community through vLLM. The result is a flywheel in which production serving on Inferact, open-source collaboration, and tuning for new workloads each feed improvements back into the others.
Why this matters
Open weights make a model available, but they do not make it operational. Serving a model at K3’s scale requires the inference engine to reproduce the model correctly, manage its attention and cache architecture, support its native capabilities, and remain reliable as traffic changes. By starting this work before release, Moonshot AI and Inferact turn open-source implementation, production qualification, and live operation into one continuous path.
For Moonshot AI, that shortens the path between a new model and dependable availability. Developers and enterprises gain a production-ready way to use Kimi models without reproducing the underlying distributed systems and performance engineering. The broader community benefits as lessons from production flow back into vLLM and strengthen the open-source serving foundation available to everyone.
What comes next
Inferact will continue to serve K3.1 Pro and future Kimi frontier models on its platform, while Moonshot AI and Inferact keep advancing vLLM support and deployment artifacts for the community. The goal is a flywheel in which production serving on Inferact, open-source collaboration through vLLM, and tuning for new workloads each feed improvements back into the others.
About Moonshot AI
Moonshot AI was founded in 2023 in pursuit of the optimal conversion from energy to intelligence. The company develops Kimi, an AI assistant for reasoning, coding, research, and agentic work, and advances frontier intelligence through open models, technical research, and co-creation with users and developers worldwide.
About Inferact
Inferact builds the production inference platform for frontier AI. Founded by the creators of vLLM, Inferact gives model builders and enterprises a continuous path from open-source model support to production serving, with deployment validation, performance engineering, capacity management, observability, and reliability across heterogeneous infrastructure.
Media contact: contact@inferact.ai
