Skip to main content

OpenCSG Community Release Notes v2026.03

1 CSGHub-Lite Officially Released​

Run LLMs on your PC with a single click. CSGHub-Lite is a lightweight tool tailored for the desktop environment. It breaks through the complexity of environment configuration, enabling non-developers to invoke AI locally securely and swiftly.

  • Ultimate Simplicity: A single binary file without complex dependencies. Simply run csghub-lite run to automatically complete model downloading, loading, and chatting.
  • Universal Format Compatibility: Built-in llama.cpp inference engine, native support for GGUF, and automatic conversion capabilities for SafeTensors.
  • Sub-second Wake-up: Innovative "resident in memory" mechanism (default 5 minutes), ensuring sub-second response for subsequent chats.
  • Resuming from Break-point: Built-in resuming from break-point mechanism, fearless of large file download interruptions; seamlessly supports macOS, Linux, and Windows.
  • App Ecosystem: Built-in common AI Apps, with the first batch including Claude Code, Codex, OpenCode, OpenClaw, etc. Supports one-click installation and uninstallation.

2 Collaboration Enhancement: Organizational Compute Sharing​

From "personal exclusive" to "team collaboration", we have broken the silos of compute resources. Now, organizations can centrally purchase or allocate compute resources, and create various model services under the organization's name.

  • Resource Democratization: Inference, Fine-tuning, Evaluation, and Notebook services all support organizational creation.
  • Efficient Collaboration: Organization members can share the same compute configuration, significantly boosting resource turnover and experimental efficiency within the enterprise.

3 Enterprise Edition (EE) Deep Upgrades: AI Infrastructure Backbone​

1. Compute Management Platform —— Panoramic Control, Compute Power at Your Fingertips A full-lifecycle compute resource management center designed for administrators, achieving automated scheduling from physical server rooms to higher-level tasks.

  • Resource Orchestration: Cluster management and standardized compute specification definition.
  • GPU Virtualization: Efficient resource multiplexing based on HAMi soft-partitioning and MIG hard-partitioning.
  • Global Observability: Health dashboards and trend tracking to monitor cluster loads in real-time.
  • Billing Analysis: GPU-Hours consumption backtracking and multi-dimensional statistical report export.

2. MCP Gateway —— Unified Tool Governance Resolves the fragmentation problem of MCP Server management within the enterprise.

  • Centralized Registration: Supports the unified access and management of remote MCP Servers.
  • Standardized Protocol: Provides a consistent access entry for upstream Agent applications via Endpoint, connection type, and Header settings.

3. Agent Sandbox —— A Runtime Environment that Balances Security and Efficiency Provides ultimate isolation and blazing-fast response sandbox capabilities for Agent execution.

  • Security Defense: Strictly prevents malicious code from attacking the host machine via system calls, ensuring complete data isolation between tenants.
  • Millisecond Startup: Achieves lightning-fast allocation and startup of the Agent environment.
  • Hot Update Support: Real-time effectiveness of code modifications without restarting the container, dramatically shortening the Agent iteration cycle.

History Version​

Click Release List to view all history versions.