EASYHUB JOURNAL / 2025-12
AI news in December 2025
11 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
HY-MT1.5 offers compact and larger translation variants
Hy-MT1.5 opens 1.8B and 7B translation models for 33 languages and five minority-language or dialect varieties. The smaller version targets quantized offline use, while the 7B update reduces unwanted commentary and mixed-language output. Developers can combine local and cloud deployment rather than treating it as a general chat model.
- 正和岛
DingTalk Agent OS brings enterprise tasks and agent tools together
DingTalk AI 1.1 brings ONE, enterprise search, the Wukong task assistant and its enterprise AI platform into the Agent OS framework, alongside hardware and industry applications. This work-and-agent platform organizes information across messages, documents, calendars and tasks; it is not a replacement desktop operating system.
- 新智元
FunctionGemma and T5Gemma 2 explore specialized compact models
The report distinguishes FunctionGemma’s compact function-calling focus from T5Gemma 2’s encoder-decoder design. Tool schemas and task adaptation still matter; neither implies unrestricted, safe control of every application without configuration.
- IT之家
Hunyuan WorldPlay opens an interactive world-model framework with spatial memory
WorldPlay, Hunyuan World Model 1.5, combines action control, spatial memory and streaming generation for first- or third-person exploration and video continuation. Tencent opens its data, training and inference framework and demonstrates revisiting geometrically consistent areas and exporting 3D point clouds, beyond static scene reconstruction.
- Alibaba Cloud
Tongyi opens compact CosyVoice3 speech generation and Fun-ASR recognition models
Tongyi releases CosyVoice3-0.5B for reference-audio-conditioned speech generation and Fun-ASR-Nano-0.8B for lightweight transcription and local fine-tuning. These models address different tasks. The broader hosted services' language coverage, latency and benchmarks should not automatically be attributed to every compact open checkpoint.
- IT之家
Nemotron 3 launches Nano first, with larger variants planned later
NVIDIA introduces a family for agent workloads, with Nano available and larger variants scheduled later. Nano’s active parameter count is different from its full stored model size, and planned releases should not be treated as already shipped.
- IT之家
LLaDA2.0 expands open discrete-diffusion language modeling
LLaDA2.0 opens MoE variants and technical material for discrete-diffusion language models. It explores a generation route distinct from conventional autoregression, with efficiency and quality dependent on tasks and implementation.
- 智东西
GLM-ASR and Xiaoao connect speech recognition to desktop input
Zhipu introduces cloud GLM-ASR and an open 1.5B Nano model, alongside Xiaoao for speech-driven text entry, translation and rewriting. The reporter also observed inconsistencies with mixed-language speech and self-corrections. Availability of a local model does not establish that every input-method feature runs offline.
- IT之家
LongCat-Image opens 6B text-to-image and image-editing models
LongCat-Image shares multiple generation and editing checkpoints focused on Chinese lettering and visual consistency. The report acknowledges gaps against some commercial systems, leaving image-by-image inspection important.
- IT之家
Tencent releases HY 2.0 Think and Instruct in Yuanbao and ima
HY 2.0 offers Think and Instruct variants, with 406B total parameters, 32B active parameters and a 256K context window for reasoning and multi-turn document work. Yuanbao, ima and cloud API access are available at announcement. Further open model and technology releases are plans, not confirmed same-day weight availability.
- IT之家
Mistral 3 includes compact 3B, 8B and 14B variants
Mistral’s family spans deployment sizes, including compact dense variants for local and lower-cost use. Results and resource requirements for the largest MoE model must remain separate from those of the smaller models.