EASYHUB JOURNAL / 2026-02
AI news in February 2026
6 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
WPS Lingxi adds GLM-5 with traceable formulas and live layout
With GLM-5 integration, WPS Lingxi generates native spreadsheet formulas such as SUMIFS and XLOOKUP, preserving references and recalculation. Canvas previews presentation edits and layout changes. Traceability here concerns inspectable formulas and document changes, not disclosure of a model’s complete internal reasoning.
- IT之家
HY-1.8B-2Bit uses quantization-aware training for smaller deployments
HY-1.8B-2Bit uses quantization-aware training on a 1.8B model, with low-bit GGUF and BF16 fake-quantized weights while retaining reasoning modes. Tencent reports roughly 300MB of weights and 600MB memory in its runtime configuration. Quantization does not change the underlying parameter count to 0.3B.
- IT之家
RynnBrain studies spatiotemporal memory for embodied planning
RynnBrain shares embodied models and evaluations around spatial memory and planning. Model-level improvements do not by themselves establish the safety or autonomy of a complete robotic system.
- IT之家
MiniCPM-o 4.5 opens full-duplex multimodal interaction research
MiniCPM-o 4.5 explores continuous perception and proactive responses with a compact model and platform adaptations. Real-time behavior depends on the complete camera, audio and inference pipeline.
- IT之家
Qwen3-Coder-Next emphasizes executable feedback for coding agents
The model uses task synthesis and environment feedback to improve longer coding workflows. Its small active parameter count reduces computation but does not reduce the full stored model to that active size.
- IT之家
GLM-OCR uses 0.9B parameters for formulas, tables and document extraction
Zhipu releases GLM-OCR and an SDK for document text, formulas, tables and information extraction, with deployment paths including vLLM, SGLang and Ollama. The 0.9B figure describes the model as a whole. Production pipelines still need document handling, and benchmark results do not imply identical accuracy on every scan or receipt.