EASYHUB JOURNAL / 2025-06
AI news in June 2025
10 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
ERNIE 4.5 opens dense and mixture-of-experts model variants
Baidu releases ERNIE 4.5 checkpoints and inference code across dense and MoE variants. Downloaded deployments differ from cloud APIs, while total parameter counts must be distinguished from active parameters.
- IT之家
Hunyuan A13B opens hybrid reasoning with 80B total parameters
Hunyuan-A13B is a hybrid-reasoning MoE model with 80B total parameters and roughly 13B active per inference. Tencent opens the model and API alongside ArtifactsBench and C3-Bench evaluation datasets. The 13B figure describes active computation, not the full weight size of a 13B dense model.
- 量子位
CodeFuse-CGM links repository graphs with patch generation
CodeFuse-CGM combines repository structure with language modeling and a Graph-RAG workflow for repair. Shared papers, code, weights and data support evaluating this alternative to more elaborate agent orchestration.
- IT之家
Confucius3-Math opens a mathematics-focused deployment option
Youdao releases a mathematics-focused model with a consumer-GPU deployment path. Vendor cost and speed comparisons are secondary to problem-level correctness and explanation quality in educational use.
- 界面新闻
Comate’s standalone IDE links design inputs, agents and live previews
Baidu’s standalone Comate AI IDE combines design-to-code, page previews and requirement clarification with the Zulu coding agent. It expands beyond the earlier code-assistant integration. Zulu’s earlier introduction and the independent IDE launch are distinct milestones; comparative performance statements in the report remain vendor claims.
- IT之家
Hunyuan3D 2.1 opens a fuller 3D asset-generation pipeline
Hunyuan 3D 2.1 improves geometry generation and opens PBR material generation with base-color, metallic, normal and roughness maps. Generated geometry and materials can enter game or rendering workflows. Local deployment documentation accompanies the release, extending the pipeline toward more complete 3D assets.
- IT之家
SmolVLA brings a 450M vision-language-action model to robotics research
SmolVLA shares a compact base model and training recipe built around public robotics data. Running inference on consumer hardware does not eliminate the need for robot hardware, task data and safe operating constraints.
- IT之家
Qwen3 opens embedding and reranking components across sizes
Qwen’s retrieval family separates embeddings from reranking and offers several sizes for multilingual and code search. The components can be combined in knowledge systems, with retrieved results still needing reliable source grounding.
- IT之家
BAGEL brings image understanding, generation and editing together
BAGEL studies unified understanding and generation through interleaved multimodal training. Released models and code support image manipulation research; navigation examples do not establish a general-purpose physical simulator.
- IT之家
Circuit Tracing opens attribution graphs for studying model behavior
Anthropic’s tracing tools and Neuronpedia interface help researchers inspect attribution graphs and example models. These graphs study internal computational relationships rather than providing a complete, infallible explanation of a model’s thinking.