Clover-35B-A3B-Preview: 35B MoE based on Qwen3.6, optimized for text-only inference

Clover-35B-A3B-Preview is a text generation model fine-tuned based on Qwen3.6-35B-A3B. It adopts a Mixture-of-Experts structure that activates only 3B out of a total 35B parameters to enhance inference efficiency. It is released under the Apache-2.0 license and designed for text-only content processing.

This model does not support image or multimodal inputs and is restricted to processing text only. The chat template allows for clear separation and management of tool calling and reasoning processes. By adjusting the preserve_thinking option, you can configure whether to retain reasoning content or preserve it only after the last user query.

It is compatible with major inference frameworks such as Transformers, vLLM, and SGLang. It provides an OpenAI-compatible API through the vLLM server, allowing for easy integration into existing LLM applications. It is stored in BF16 precision and distributed in a sharded file format to support loading in large-scale GPU environments.

It is suitable for text-based conversational AI or agent systems requiring tool calling. It is useful for developers seeking lightweight inference performance in pure text tasks where multimodal capabilities are unnecessary. As a Preview version, it is appropriate to approach it for experimental purposes before the final stable release.