DiffuSpace has open-sourced Clover-35B-A3B-Preview, a 35B-parameter diffusion language model with roughly 3B active parameters that uses block-wise diffusion for parallel token generation, reaching 5-10x the throughput of comparable autoregressive models. Under a unified evaluation protocol it tops all seven agent benchmarks reported, scoring 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1, with the technical report and SFT code released alongside. The team expects diffusion LMs to replace same-scale autoregressive models in some scenarios within one to two years.
Autoregressive Model
自回归模型
A generative architecture that predicts subsequent sequence elements based on prior outputs.
Curated Intelligence
DiffuSpace has secured hundreds of millions of RMB in two consecutive funding rounds, setting a global record for diffusion language model (dLLM) financing. Its flagship model, Dream 7B, reportedly outperforms autoregressive models of similar size and matches the planning capabilities of the 671B-parameter DeepSeek V3. With major players like Google and Ant Group accelerating their efforts, dLLMs are rapidly transitioning from frontier research to engineering validation.
To address high inference costs in complex AI search caused by query decomposition, the 'Retrieve-for-Train' method is proposed. It replaces zero-shot LLM dynamic planning with retrieval-augmented training, significantly reducing reliance on expensive thinking budgets for efficient set retrieval.
The vLLM-Omni team detailed engineering solutions for adapting multi-modal speech models like Qwen3-TTS and VoxCPM2, addressing the conflict between latency-bound Talker stages and throughput-bound Code2Wav modules. By decoupling scheduling strategies and optimizing streaming pipelines, the approach enhances serving efficiency for multi-stage autoregressive models, offering a reusable architectural reference for open-source TTS deployment.
Zhipu AI and Huawei have open-sourced GLM-Image, a new multimodal model fully trained on Ascend Atlas 800T A2 chips using the MindSpore framework, validating domestic hardware for frontier AI. The model employs a hybrid 'autoregressive + diffusion decoder' architecture, achieving state-of-the-art performance in complex text rendering (especially Chinese characters) with ultra-low inference costs.
AI expert Sebastian Raschka published an article surveying alternative LLM architectures beyond standard autoregressive Transformers. The piece covers emerging directions like text diffusion models, linear attention hybrids, and code world models, addressing community interest in non-mainstream approaches.
OpenAI researchers propose a data generation pipeline combining domain randomization with autoregressive models for robotic grasping. By training on millions of procedurally generated unrealistic objects, the model achieves over 90% success rate on unseen realistic objects in simulation, addressing generalization challenges caused by limited training data diversity.