Run Kimi K3 in the EU through an OpenAI-compatible API

Kimi K3 launched in July 2026, with full weights published on 27 July. It is a mixture-of-experts model with 2.8 trillion parameters, 104 billion of them active per token: each token is routed to 16 of 896 experts. It has 93 layers, most of which use Moonshot's Kimi Delta Attention, mixed with gated latent attention layers.

Moonshot says the new architecture gives about 2.5 times the scaling efficiency of Kimi K2. The model reads text, images and video natively and has a one-million-token context window. It always reasons, with effort levels low, high and max, and was trained with quantisation-aware training in MXFP4 weights. It is released under its own Kimi K3 licence.

Kimi K3 is aimed at the hardest open-model workloads: long engineering sessions on large repositories, deep research with many tool calls, and knowledge work that mixes documents, images and video. Its one-million-token window lets an agent keep huge amounts of material in view. For tasks where a mid-sized model keeps failing, K3 is a strong escalation step.

It is an extremely large model, so it costs more per token than almost anything else in our catalogue, and its default max reasoning produces long outputs. For everyday tasks, Kimi K2.6, GLM-5.3-Flash or DeepSeek V4.1 Flash are more economical. Its training reflects its maker's choices on sensitive topics, and its licence has conditions worth reading.

inference.eu offers the OpenAI and the Anthropic request formats. Set the base URL to the inference.eu API, use your inference.eu key and set the model to kimi-k3. Coding agents and research tools that accept a custom endpoint work without further changes.

Kimi K3 always thinks and returns reasoning_content alongside the answer. Set reasoning_effort to low, high or max; the default is max. K3 was trained with preserved thinking history, so in multi-turn conversations and tool loops pass each full assistant message back unchanged, including its reasoning. Send images as image_url parts. Moonshot uses temperature 1.0 in its evaluations.

Kimi K3 is developed in China, but on inference.eu the open weights run on infrastructure in the European Union. Your prompts, images, video and the model's responses are processed in EU data centres, are not sent to Moonshot AI, and are never used to train any model.

Every customer gets a data processing agreement under Article 28 of the GDPR and our subprocessor list. Billing is in euros. The Kimi K3 licence permits use, modification and sale. Products with more than 100 million monthly active users or 20 million US dollars of monthly revenue must display "Kimi K3" prominently, and internal use that exposes nothing to third parties is exempt from the extra conditions.