DiffuSpace Open-Sources Clover-35B-A3B-Preview Diffusion Language Model
AI lab DiffuSpace has released Clover-35B-A3B-Preview, the first model in the Clover series, built on a MoE architecture and open-sourced under the Apache license. The model generates results through multiple rounds of iterative generation; DiffuSpace claims it outperforms the same-scale autoregressive model Qwen3.6-35B-A3B on multiple benchmarks while maintaining faster generation speeds, and is particularly strong at agentic coding and tool-use tasks.
Why read this
AI score 68
The original text presents a comparison between a diffusion language model at the 35B-parameter MoE scale and autoregressive models of the same scale, which can serve as a reference for judging the feasibility of this technical approach.
Topics
Body · AI translation
DiffuSpace Launches Clover-35B-A3B-Preview Diffusion Language Model, Balancing Performance and Speed
English text is being prepared.
English text is being prepared.
Clover-35B-A3B-Preview is a diffusion language model using a MoE architecture. Unlike traditional autoregressive models, it possesses "global thinking," achieving results from vague to concrete through multiple rounds of iteration.
DiffuSpace stated that the birth of Clover-35B-A3B-Preview proves that diffusion language models can compete with, or even slightly outperform, autoregressive models at larger parameter scales while maintaining faster generation speeds.
The model outperforms the same-scale autoregressive model Qwen3.6-35B-A3B on multiple benchmarks, and is especially adept at agentic coding and tool-use tasks.
Reference
English text is being prepared.