EASYHUB JOURNAL / 2026-04
AI news in April 2026
11 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
ima copilot adds editable memory and knowledge-library skills
ima copilot works with the page, file or note currently open and retains context through user profiles, long-term memory and experience records. Users can inspect and edit memory and invoke or extend library, note and report skills. The multi-platform rollout initially requires an access application.
- IT之家
Hunyuan’s Android demo applies low-bit models to offline translation
Tencent compresses Hy-MT1.5-1.8B using 2-bit and 1.25-bit schemes, with the latter’s weights around 440MB, and provides an Android offline translation demo. The 33-language model supports local translation with background text capture. The file size is not total runtime memory; iOS support remains planned at announcement.
- IT之家
QoderWork expert packs combine domain knowledge, workflows and connectors
QoderWork introduces ten expert packs across six domains, combining reference knowledge, workflows and data connectors. Users can start research, contract-document or marketing-planning tasks without writing a skill from scratch, connecting information retrieval, structured analysis and report drafting in one reusable process.
- IT之家
QClaw v0.2.14 adds Hermes, model selection and an expert gallery
QClaw v0.2.14 supports Hermes alongside its existing agent engine and introduces a scenario-based expert gallery. It adds model selection, voice and file sharing in the WeChat mini program, and management of linked cloud instances. Usage points depend on tasks and models rather than unlimited expert access.
- IT之家
Privacy Filter opens a model for detecting personal information in text
The report covers an open model for identifying personal information in text and lower-resource deployment. False positives and missed sensitive spans remain possible, so it should complement rules and review rather than replace governance.
- IT之家
Hy3 preview opens and reaches Tencent’s office and development products
Hy3 preview combines fast and deliberative modes in a MoE model with 295B total parameters, 21B active parameters and 256K context. Tencent integrates it into Yuanbao, CodeBuddy, WorkBuddy and ima, opens cloud access, and supports agent tools including OpenClaw and OpenCode for coding and multi-step work.
- IT之家
ERNIE-Image opens an 8B image model and ComfyUI workflow
ERNIE-Image pairs an image model with prompt enhancement and deployment assets for controlled layouts. Consumer-GPU demonstrations are configuration-specific, while complex lettering and composition still need review.
- IT之家
MiniMax Agent adds Pocket remote access and computer-use capabilities
MiniMax Agent’s Pocket Beta accepts desktop-task requests from Feishu, WeChat, WeCom or Slack and sends results back to the conversation. The same update adds computer use through screen interpretation, mouse and keyboard actions, extending automation to local software workflows not covered by files and APIs alone.
- IT之家
MiniMax MMX-CLI gives agents a consistent interface to multimodal models
MMX-CLI provides a consistent command interface for agents such as Claude Code and OpenClaw to use MiniMax coding, video, speech and music models. It separates progress messages from machine-readable results and supplies categorized exit codes for recovery. No custom MCP server is required, but model calls still use service allowances.
- IT之家
QClaw V2 adds three parallel agents and cross-application connectors
QClaw V2, version 0.2.5, supports up to three parallel agents with separate roles and conversations. Initial connectors include Tencent Docs, Meetings, ima, Notion and email, allowing authorized tasks to write results into other applications. A new protection manager is also introduced, without making all automation risk-free.
- IT之家
Gemma 4 combines edge-oriented and larger open model variants
Gemma 4 spans edge-oriented and larger variants with tool-use and multimodal capabilities. Effective runtime parameters differ from total model scale, and hardware and modality support vary by version.