EASYHUB JOURNAL / 2025-04
AI news in April 2025
10 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
Xiaomi MiMo opens a 7B reasoning model and technical report
Xiaomi introduces MiMo through its official account, linking pretraining and post-training for mathematics and coding. Reported comparisons concern particular evaluations rather than every use case for larger models.
- IT之家
Step1X-Edit opens natural-language-controlled image editing
Step1X-Edit targets text replacement, style changes and region-specific image edits. Identity preservation and precision are evaluation goals, not guarantees; important portraits and lettering still need manual review.
- Moonshot AI / Kimi
Kimi-Audio opens a 7B instruction checkpoint and audio evaluation tools
Kimi-Audio combines transcription, audio questions and spoken conversation, initially releasing a 7B instruction checkpoint, inference code and an evaluation toolkit. Developers can extract text from recordings or generate paired speech and text responses, using a shared interface for audio understanding and conversational-assistant experiments.
- IT之家
SkyReels-V2 opens research into longer video generation
SkyReels-V2 explores longer and continued video generation using diffusion-forcing methods. Claims of unlimited duration do not imply unlimited compute or consistent quality, so longer outputs still need resource and continuity management.
- IT之家
BitNet b1.58 2B4T explores native low-bit language modeling
BitNet studies native low-bit training rather than only post-training compression. Its efficiency depends on a supporting runtime, and non-embedding weight memory should not be mistaken for total application memory.
- IT之家
UI-TARS-1.5 improves visual interface actions through reinforcement learning
UI-TARS-1.5 improves visual perception, planning and cross-platform actions. Browser and device demonstrations support evaluation, while real automation still needs constrained permissions and checked outcomes.
- IT之家
Codex CLI opens a terminal-based coding-agent interface
Codex CLI connects model interactions to local files and terminal commands. Open-source client code does not make model usage free, and command permissions, scope and version control remain separate responsibilities.
- IT之家
InternVL3 expands multimodal models and interface understanding
InternVL3 spans multiple parameter scales and explores interfaces, drawings and spatial reasoning. A shared family release does not imply identical capabilities or hardware requirements across its variants.
- IT之家
Skywork-OR1 opens reasoning models with training data and code
Skywork-OR1 publishes reasoning checkpoints alongside training assets and methodology. The materials enable community experimentation, while reproduced results depend on hardware, data versions and configuration.
- The Verge
Canva Visual Suite 2.0 combines design, spreadsheets and interactive content
Canva combines design and productivity tools with coding, image editing and spreadsheets. Generated interactive elements still need testing before deployment, and data-derived outputs require independent checking.