New Chinese AI Agent Breaks TerminalBench and Destroys Claude Opus 4.6

AI Developments: CodeBrain, Cedance 2.0, Quen Image 2.0 & Fine-Grained Recognition

Key Concepts:

• Terminal Bench 2.0: A benchmark for evaluating AI agent performance in real-world computer tasks.
• Language Server Protocol (LSP): Tools enabling AI agents to access and utilize specific code and documentation.
• Membrane: Feeling AI’s long-term memory solution for AI agents.
• Cedance 2.0: Bite Dance’s new AI video model focused on intentional and consistent scene generation.
• Quen Image 2.0: Alibaba’s image generation model emphasizing prompt adherence and creative control.
• Fine-Grained Recognition: The ability of AI to distinguish between highly similar objects (e.g., different aircraft types).
• Fine R1: A vision model developed by PKing University specializing in fine-grained recognition with minimal training data.
• Token: A unit of data used by AI models; fewer tokens generally mean faster and cheaper processing.

AI Agents: CodeBrain 1’s Breakthrough Performance

The AI agent landscape has seen a significant shift with the emergence of CodeBrain 1, developed by Feeling AI. CodeBrain 1 achieved a score of 72.9% on the Terminal Bench 2.0, placing it second globally, a remarkable feat for a first-time entrant. This benchmark, increasingly recognized as a crucial “stress test” for AI agents, assesses their ability to perform tasks within a computer environment, rather than simply discussing them.

OpenAI currently leads with scores of 77.3% and 75.1% (using GPT-5.3 codeex and a related setup), followed by Anthropic’s Claude Opus 4.6 at 65.4%. CodeBrain 1’s 72.9% (with another configuration at 70.3%) surpasses many other leading labs and agent systems.

The success of CodeBrain 1 is attributed to its focus on practical code execution. Instead of relying on broad generalizations, it prioritizes ensuring code runs correctly by utilizing the Language Server Protocol (LSP) to access precise code references and documentation relevant to the task at hand. For example, when tasked with programming a gamebot, it directly references function names like “move to target” and “do action” rather than attempting to infer them.

Furthermore, CodeBrain 1 excels in error handling. When code fails, it analyzes diagnostics, examines correctly written similar code, and consults documentation to pinpoint and rectify the issue. Testing on 47 Python tasks demonstrated system stability and a 15% reduction in token usage compared to some competitors, resulting in lower costs and faster processing times.

CodeBrain 1 also exhibits dynamic planning capabilities. It breaks down complex requests into sequential steps (e.g., “Gather resources, clear space, craft tools, build structure” for building a house in a game) and adapts its strategy based on experience, even enabling “group memory” in tactical games where enemy AI adjusts to player behavior. This is complemented by Feeling AI’s Membrane technology, which provides long-term memory capabilities, achieving over 300% improvement on the challenging nomi bench level 3 benchmark. Combined, Membrane and CodeBrain 1 represent advancements in both memory and planning for AI agents.

AI Video: Cedance 2.0 and the Shift Towards Intentional Content Creation

Bite Dance’s Cedance 2.0 is being hailed as a potential turning point in AI video generation. Unlike previous models that produced visually chaotic and inconsistent results, Cedance 2.0 demonstrates an understanding of scene flow and delivers more intentional and consistent video output. It accepts multimodal inputs – text, images, video, and audio – allowing for versatile control.

Bite Dance is aggressively promoting adoption by offering access for just one UNO (with auto-renewal) on their gyming platform. Cedance 2.0 functions more like a “digital director,” handling camera motion (push-ins, pans, tilts, tracking shots) in a way that feels planned and deliberate. Crucially, it maintains character consistency and scene stability, enabling the creation of longer, story-driven clips.

Industry expert Fang G from game science predicts this will lead to “content inflation” as the cost of video production decreases dramatically. This will particularly impact e-commerce (reducing the need for expensive product videos) and gaming (accelerating the creation of trailers and promotional material). For platforms like TikTok, the challenge will shift from content creation to content filtering.

The shift also impacts traditional film and TV production, moving the workflow from shooting and editing to “describe the scene, generate the scene,” with editors evolving into creative directors guiding the AI tool. However, this raises copyright concerns, exemplified by Steven Chow’s team questioning the legality of AI-generated parodies mimicking his style. The speed of development is remarkable, with Cedance 2.0 achieving 60-second, audio-driven narrative videos with multimodal inputs just two years after OpenAI’s Sora 1.0.

AI Image Generation: Quen Image 2.0 and Enhanced Prompt Following

Alibaba’s Quen Image 2.0 addresses a common frustration with AI image generation: the tendency to ignore parts of detailed prompts. Quen Image 2.0 can process up to 1,000 tokens of instructions, follow complex scene descriptions, accurately render Chinese text, edit existing images, and output images up to 2K resolution.

Demonstrations included generating a five-panel ink-style comic based on Journey to the West with consistent characters and distinct environments, a detailed hamburger infographic with realistic textures, and a Shanghai city scene combining multiple artistic styles. The model also excels at editing, seamlessly combining images and transforming selfies into professional-looking photos. Its ability to accurately render complex Chinese text, such as the preface to the Orchid Pavilion collection, represents a significant improvement over previous models.

International tests position Quen Image 2.0 just behind Nano Banana Pro, placing it among the top image generation models globally.

Fine-Grained Recognition: Fine R1 and Minimal Data Learning

Researchers at PKing University have developed Fine R1, a vision model capable of distinguishing between highly similar objects, such as different aircraft types, using a remarkably small amount of training data. This is a challenging task known as “fine-grained recognition.”

Fine R1 employs a structured visual reasoning approach, analyzing visual details, listing possible subcategories, comparing them, and then making a decision. It achieved superior performance to models like Clip and Sigpip on several fine-grained datasets, even with only four training images per category, and can directly name categories without being provided options. The training method involves presenting the model with images from the same subcategory, and a very similar image from a different subcategory, forcing it to learn subtle differentiating details.

Conclusion:

The advancements across AI agents (CodeBrain 1), video generation (Cedance 2.0), image generation (Quen Image 2.0), and visual recognition (Fine R1) demonstrate a clear trend towards more structured, reliable, and capable AI systems. These developments are not only pushing the boundaries of what’s technically possible but also reshaping industries and raising important questions about copyright and the future of content creation. The focus on practical application, efficient resource utilization, and the ability to handle complex tasks with minimal data are key takeaways from these recent breakthroughs.