Since the release of KAT-Coder-Pro V1, we have continuously received valuable feedback and suggestions from frontline developers. These real-world usage insights have driven us to constantly refine and expand the capability boundaries of the KAT series models in practical application scenarios.
KAT-Coder-Pro V2 is the latest flagship Agentic Coding model developed by KwaiKAT. In Agentic scenarios, KAT-Coder-Pro V2 features powerful scaffolding generalization capabilities, compatible with more than 10 mainstream AI coding tools such as Claude Code, Cline, Kilo, and OpenCode, providing greater flexibility. It has also undergone specialized training and deep optimization for OpenClaw, enabling it to handle complex real-world application workflows with ease.
Meanwhile, KAT-Coder-Pro V2 has achieved breakthrough progress in frontend aesthetic generation—in Landing Page and PPT scenarios, users only need to provide colloquial descriptions to obtain high-quality outputs approaching the level of structured design spec inputs. This means the model's service boundary has truly expanded from the past 1% of professional users to hundreds of millions of ordinary users.
Native Adaptation to OpenClaw, Deep Optimization for Multi-Agent Frameworks
In real-world AI Coding deployment scenarios, AI Agent frameworks represented by OpenClaw continue to iterate at high frequency, constantly introducing new tools and protocols, which poses tremendous challenges to the model's scaffolding generalization capabilities. Tool call failures, multi-step task interruptions, and instruction understanding deviations are problems frequently exposed during actual use, and in high-frequency usage scenarios, these issues are magnified exponentially, directly impacting user experience.
The true boundary of model capabilities lies not only in whether code generation quality passes muster, but also in whether it can accurately understand user intent throughout long-horizon trajectories when facing complex environments where tools constantly expand and task chains continue to lengthen, and maintain stable, consistent performance across different Agent frameworks. Whether it's Claude Code or OpenClaw, users should be able to switch seamlessly and use it with confidence, rather than stumbling through new pitfalls when switching frameworks.
To this end, KAT-Coder has systematically designed around multi-scaffolding generalization capabilities from data construction to training processes, and performed full-chain specialized optimization for OpenClaw usage scenarios based on native task data—covering not only scaffolding protocol understanding and tool chain invocation, but also deeply reinforcing long-link execution stability during the training phase.
Final evaluation results show that KAT-Coder has achieved significant improvements in complex Skills compliance rates and multi-step task completion rates, with execution efficiency and response stability in high-pressure scenarios such as scheduled triggers, high throughput, and long linkages simultaneously reaching world-class standards.
It is worth mentioning that KAT-Coder-Pro V2's scaffolding generalization capability is not limited to the single OpenClaw framework. We have simultaneously conducted evaluations on mainstream scaffolds such as Claude Code and OpenCode, and results show that the model also possesses excellent adaptation capabilities in cross-framework scenarios.
Web Coding - When Models Start Understanding "Aesthetics"
Breaking Old Consensus: Systematic Blind Spots in Existing Evaluations
Current mainstream code generation evaluations (such as WebArena, etc.) essentially play "spot the difference"—given a reference image, see how accurately AI can copy the cat and draw the tiger. But this creates serious misalignment in "one-sentence generates webpage" scenarios.
In commercial applications: "Code runs" and "Design looks good" are two different things. Code fidelity measures "is the code correct" (are there errors or misalignments), which can be calculated by algorithms; Aesthetic fidelity measures "does the page look good", which is advanced aesthetic judgment—code running is just the starting line.
Existing evaluation standards are seriously biased, leaving six major blind spots:
1. Users only give one sentence, there is no "standard answer" for AI to compare against.
2. Image algorithms will give low scores to breakthrough original designs.
3. Static screenshots cannot capture the quality of interaction animations.
4. Algorithms cannot quantify abstract terms like "high-end business feel".
5. Only looking at whether individual buttons are good, regardless of whether overall layout matches.
6. Existing algorithm scoring actually forces AI toward the most mediocre, safest designs.
KAT Benchmark: A New Industry Benchmark Based on Professional Design
Based on the profound humanistic vision and frontend foundation of Kuaishou's R&D design team, we have filled the gap and launched the "KAT Aesthetic Benchmark" calibrated by professional designers and cooperative teams.
As the industry's only pure aesthetic standard for "creation without reference images", it has four major advantages:
• Insists on designer manual blind testing, rejecting algorithm-onlyism—true aesthetics cannot be replaced by machines.
• Pioneers 10 independent evaluation dimensions, with granularity far exceeding existing academic standards.
• "Outstanding without flaws" is the full score, not "most similar to reference image", encouraging originality and punishing mediocrity.
• Rigorous design and review mechanism, professional designer teams perform deep interaction blind testing under unified standard screens.
The Data Speaks
Under the strictest ruler, KAT shows dominance:
• PPT scene crushing: Total score 57.6, leading competitors by 14–22 points, color scheme item up to 78 points; image scores are 5–8 times those of competitors.
• Landing Page topping the charts: Total score 59.8 takes first place, establishing insurmountable advantages in color scheme, elements, and layout.
• Shocking leap: Compared to last generation baseline, PPT average score doubled (+103%), LP improved +42%, individual element item surged +300%.
Every leap on the Benchmark is making "one sentence generates professional-level commercial pages" step by step toward reality.
PPT Case
Landing Page Case
Stronger Base Capabilities
Complex reasoning in Agentic scenarios cannot be separated from solid general base capabilities as support. KAT-Coder-Pro V2's base model has entered the global first tier on mainstream benchmarks such as Terminal-Bench Hard (46.8), τ²-Bench Telecom (93.9), providing solid underlying guarantees for upper-layer Coding capabilities.
Get Started Now
KAT-Coder-Pro V2 is now fully online. Users can immediately experience it through the following methods:
Method One: API Call
Call the model API directly through the StreamLake.com platform and flexibly integrate it into your workflow.
API Key Application:
https://streamlake.com/product/kat-coder
Method Two: Coding Plan Subscription
KAT-Coder-Pro V2 has been included in the Coding Plan package, ready to use out of the box. We provide four tiers of plans, and you can choose according to your usage frequency:
Coding Plan Subscription:
https://www.streamlake.com/marketing/coding-plan
Development Tool Integration Guide:
https://www.streamlake.com/document/WANQING/me6ymdjrqv8lp4iq0o9