Get top-tier performance at half the price! Try the TianGong SkyClaw Agent model for free for a limited time.
On May 26th, TianGong AI under Kunlun Wanwei officially launched the high-performance Agent model SkyClaw-v1.0 and simultaneously released the lightweight version SkyClaw-v1.0-lite, which combines top-notch performance with extreme cost-effectiveness.
SkyClaw-v1.0 supports a context of one million tokens and is deeply adapted to various real-world intelligent agent work scenarios. It focuses on optimizing complex tool calls, multi-round task execution, code generation, file editing, interactive application construction, and research-oriented data analysis. The model has been optimized through large-scale mid-training, high-quality synthetic task SFT, and end-to-end reinforcement learning. It can run in mainstream Agent environments such as OpenClaw, Hermes, and Nanobot, and is also compatible with code Agent frameworks such as Claude Code and Codex, maximizing its versatility and compatibility.
The combination of the Agent model and the Agent framework is changing the way models are used. In the past, models mainly answered questions; now, models are placed in an execution environment that can read repositories, call tools, edit files, run tests, and observe feedback, and start to undertake complete workflows. SkyClaw-v1.0 is designed for this stage: enabling the model to continuously advance tasks in a long-context and tool environment, rather than just generating a single answer.
Core Highlights of the Model
Strong Task Completion Ability: In mainstream Agent benchmarks and internal Claw task evaluations at Skywork, SkyClaw-v1.0 demonstrates stable multi-step task execution ability.
Comprehensively Surpasses Mainstream Open-Source Models: The model outperforms Minimax 2.7, DeepSeek V4 Flash, as well as Qwen 3.6 35B A3B and 27B models.
Approaches Larger-Scale Top-Level Models: In OpenClaw-related tasks, SkyClaw-v1.0 performs similarly to larger-scale models such as DeepSeek V4 Pro, Claude Opus 4.6, and Qwen 3.6 Plus.
Highly Cost-Effective: The pricing is less than half of that of Minimax 2.7 and Qwen 3.6 series models, providing a foundation for large-scale invocation of high-performance Agent capabilities.
For application construction tasks, we recommend running SkyClaw-v1.0 in Agent frameworks such as Hermes, Claude Code, or Codex, enabling it to complete planning, file editing, test execution, and multi-round iterations, rather than just generating code snippets.
Core Training Concept: Focus on Real-World Task Fulfillment Ability
The training goal of SkyClaw-v1.0 is clear: to improve the model's ability to complete real tasks in the Agent framework. The training focuses on three aspects: building an interactive tool environment, screening high-quality task trajectories, and using reinforcement learning to enhance the stability of multi-step execution.
Agent Environment Setup
The training environment is built based on OpenClaw-style agent frames, covering high-frequency Agent actions such as file reading, code editing, retrieval, testing, and page observation. During training, the model not only generates answers but also needs to select and combine tools and continue to advance tasks based on the tool return results.
The team further combines real Claw task data and online skill usage feedback to build a tool relationship graph, which is used to synthesize complex tasks closer to real workflows. The data obtained in this way is not isolated Q&A but a complete execution chain that includes goal decomposition, tool invocation, result observation, and iterative correction.
Refined Synthetic Training Data
SkyClaw-v1.0 uses a large amount of synthetic Agent trajectories for mid-training and SFT.
This step mainly addresses the noise problem in Agent training. Low-quality trajectories can cause the model to learn ineffective tool calls, incorrect observation interpretations, or intermediate steps that deviate from the goal. SkyClaw-v1.0 retains more stable and reusable task execution patterns through trajectory quality filtering and data ratio experiments.
Agentic Reinforcement Learning Iteration
The reinforcement learning stage continues in the self-built Claw environment. The model needs to execute tasks, observe feedback, handle failures, and continue to correct actions in the interactive environment. The optimization goal shifts from "whether the answer looks good" to "whether the task is completed and whether the process is stable."
Therefore, when SkyClaw-v1.0 is used in environments such as OpenClaw, Hermes, Nanobot, Claude Code, and Codex, its advantages are more reflected in continuous execution, error recovery, and multi-round iterations, rather than the superficial completeness of a single answer.
Comprehensive Verification of Capabilities in Multi-Scenario Practical Applications
In the Agent framework, SkyClaw-v1.0 can complete planning, file editing, code generation, test running, page debugging, and multi-round iterations. It is more suitable for delivering complete applications, interactive games, and research-oriented web reports. These examples all start from natural language prompts and are completed in Agent frameworks such as Hermes, Claude Code, and Codex. We strongly recommend using SkyClaw-v1.0 as a model in the Agent workflow rather than as an independent chat model.
Full-Form Interface Design Implementation
SkyClaw-v1.0 can generate application interfaces with production-level layouts, real navigation processes, and complete interaction states, covering common product forms such as multi-page structures, list filtering, detail pages, forms, and mobile adaptation.
b. Immersive Interactive Game Development
SkyClaw-v1.0 can generate runnable interactive games and physical simulations. It not only outputs page structures but also can handle animation loops, collision detection, game rules, state management, and user input.
c. Professional Web Page Production and In-Depth Research Analysis
SkyClaw-v1.0 is also suitable for generating research-oriented web pages and data reports. The model can organize information, sort data, design pages, and present visualizations around open topics, converting natural language requirements into interactive web reports.
Now Integrated into TianGong Skywork and Open for Free Trial
The SkyClaw-v1.0 model was integrated into TianGong Skywork on May 22nd, 2026. Users can log in to https://tiangong.cn, open TianGong Skywork, and use it directly without additional configuration of the Agent environment.
As of now, SkyClaw-v1.0 and SkyClaw-v1.0-lite are open for a free trial for 2 to 4 weeks. Users can experience the model's long-context understanding, multi-round execution, code generation, tool invocation, and application construction capabilities in real tasks.
Project Address: https://skyworkai.github.io/skyclaw/
API Access Address:
SkyClaw-v1:
https://www.apifree.ai/model/skywork-ai/skyclaw-v1?tab=api
SkyClaw-v1-lite:
https://www.apifree.ai/model/skywork-ai/skyclaw-v1-lite?tab=api
For developers, SkyClaw-v1.0 also provides free API calls through APIFree. After registering an APIFree account and obtaining an API Key, you can access existing applications or Agent frameworks through an interface compatible with the OpenAI format. The API supports streaming output, tool invocation, and multi-round conversations, making it suitable for accessing code Agents, self-developed workflow systems, enterprise internal tools, and automated task platforms.
The launch of the SkyClaw-v1.0 model is not an isolated model release but a key part of Kunlun Wanwei's AGI product system. In the product system, SkyClaw-v1.0 undertakes the upgrade of the underlying Agent model capabilities: through one million context, Agentic RL, optimization of complex tool calls, and cost-effective APIs, it advances Agents from the "demonstrable" stage to the stage of "high-frequency invocation and real delivery."