← all releases
Workshop Labs·Mar 12, 2026·v1.0·open source
Trellis-KimiK2T
Trellis-KimiK2T is an open-source post-training codebase developed by Workshop Labs, designed to enable rapid LoRA training on Kimi-K2-Thinking models. It achieves a training speed of 6,600 tokens per second on a single 8xH200 node, making it 50 times faster than the best open-source alternative and more than twice as cost-effective as the closest private training API.
Agentic AIInfrastructure
overview
Trellis-KimiK2T is an open-source post-training codebase developed by Workshop Labs, designed to enable rapid LoRA training on Kimi-K2-Thinking models. It achieves a training speed of 6,600 tokens per second on a single 8xH200 node, making it 50 times faster than the best open-source alternative and more than twice as cost-effective as the closest private training API.
key features
- 01Open-source codebase
- 02Rapid LoRA training
- 03High training speed (6,600 tokens/sec)
- 04Single-node implementation
- 05Cost-effective compared to alternatives
related productsbrowse all →
AIO Sandboxopen sourceAIO Sandbox is an integrated environment designed for AI agents, combining a browser, terminal, filesystem, VSCode, Jupyter, and MCP Server into a single Docker container. This unified setup allows seamless development and execution of AI agents without the need for multiple services.Mistral Forgeopen sourceMistral Forge is a system designed for enterprises to build frontier-grade AI models grounded in their proprietary knowledge. It enables organizations to train models on internal data, ensuring alignment with unique operations and compliance requirements.Fractalsopen sourceFractals is an open-source recursive agentic task orchestrator developed by TinyAGI. It enables the decomposition of high-level tasks into self-similar trees of executable subtasks, which are then executed by an agent swarm in isolated git worktrees.