China’s New Open AI Shocks OpenAI: DESTROYS Closed Model Limits (Better Than DeepSeek & Kimi)

GLM 4.7 & Manis Design View: A Detailed Overview

Key Concepts:

• GLM 4.7: An open-source large language model (LLM) focused on coding and agent-based workflows, demonstrating significant performance improvements in coding benchmarks, terminal tasks, and reasoning with tools.
• Agent-Friendly Systems: LLMs designed for complex, multi-step tasks requiring planning, tool use, and consistent execution over extended periods.
• Preserved Thinking: A GLM 4.7 feature that maintains internal reasoning state across multiple turns, reducing drift and improving long-horizon task stability.
• Manis Design View: A new feature within the Manis platform enabling precise, localized editing of AI-generated images, moving beyond simple regeneration.
• Nano Banana Pro: Google’s high-fidelity image model powering Manis’ photorealistic generation and editing capabilities.
• Iterative Workflow: The shift in AI image editing from prompt-regenerate cycles to a generate-refine approach facilitated by Manis Design View.

GLM 4.7: A Leap for Open-Source Coding Agents

GLM 4.7, the latest iteration in Japu’s GLM series, is positioned as a “coding-first, agent-friendly” system. This signifies a deliberate design choice prioritizing sustained performance in complex workflows over short-burst conversational cleverness. The core challenge addressed is the tendency of LLMs to “lose the thread” during extended tasks, forgetting earlier changes and becoming inconsistent. GLM 4.7 aims to overcome this through substantial performance gains across key benchmarks.

Performance Benchmarks:

• SWEBench Verified: 73.8% – A significant milestone for an open-source model, demonstrating ability to understand, modify, and integrate code within existing projects.
• Live Codebench v6: 84.9% – Reflects performance on real-world coding tasks involving constraints, edge cases, and logical reasoning.
• SWEBench Multilingual: 66.7% – A substantial improvement over GLM 4.6, highlighting enhanced multilingual coding capabilities.
• Terminal Bench 2.0: 41% – A major increase, indicating improved reliability in executing sequential commands and managing state in terminal environments. Performance in the “hard terminal” benchmark is in the low 30s, with the key takeaway being the consistent improvement across the terminal suite.
• Humanity’s Last Exam (with tools): 42.8% – A significant jump compared to performance without tools, demonstrating the model’s ability to leverage external capabilities.
• MMLURO, GPQA Diamond, Math/Competition Benchmarks: Consistent gains across various reasoning-heavy benchmarks.

Key Features & Methodologies:

GLM 4.7 introduces three “thinking modes” designed to enhance agent stability:

• Interled Thinking: Reasoning before each response or tool call.
• Preserved Thinking: Maintaining reasoning state across multiple turns, reducing drift and cost. As stated, “Preserved thinking reduces that drift because the model can carry forward its internal reasoning state across turns.”
• Turn-Level Thinking Control: Adjustable reasoning intensity based on task complexity.

The “preserved thinking” mode is particularly noteworthy, as it addresses the common issue of LLMs losing coherence over extended sessions due to reasoning resets.

Real-World Integration & Ecosystem:

GLM 4.7 is integrated with:

• Z.AI: Via an API platform supporting standard and streaming usage.
• Open Router: Providing global access and integration into existing stacks.
• Compatible with Coding Agents: Specifically mentioned as compatible with Claude Code, Klein, Rue Code, Kilo Code, and Trey-style setups.

The model also boasts high token throughput on specialized inference providers, indicating a focus on practical deployment.

Limitations:

While GLM 4.7 represents a significant advancement, top proprietary models can still outperform it on certain complex tasks, particularly in zero-shot scenarios. Deployment requires substantial hardware, even with quantized variants. However, the model is positioned as a cost-effective alternative to premium proprietary options, making agent workflows more accessible.

Manis Design View: Redefining AI Image Editing

Manis has launched “Design View,” a feature designed to transform AI image generation from a “lottery machine” into an editable workflow. The traditional AI image generation process – prompt, generate, regenerate – is often frustrating due to the difficulty of making precise edits without altering the entire image.

Core Functionality:

• Precise Selection Tools (“Mark Tool”): Allows users to highlight specific regions for editing.
• Localized Edits: Changes are applied to the selected area while preserving the characteristics of the rest of the image.
• Editable Text Overlays: Enables clean, editable text within generated images, addressing a common AI image generation weakness.
• Element-Level Editing for Slides: Allows modification of individual elements within AI-generated slides, including text, visuals, and formatting.
• Bulk Edits: Facilitates consistent changes across multiple slides, streamlining presentation workflows.

Underlying Technology:

Manis leverages Google’s Nano Banana Pro for photorealistic image generation and editing. The key is Nano Banana Pro’s ability to maintain consistency during localized edits, preventing the image from appearing “patched together.” Manis emphasizes preserving “original image characteristics” during edits, ensuring a cohesive visual result.

Comparison to Existing Tools:

Manis differentiates itself from tools like Photoshop’s Generative Fill and Canva’s Magic Edit through its integrated workflow. While Photoshop offers powerful editing capabilities, it’s a separate environment. Canva provides accessibility but with different editing logic. Manis aims to be an all-in-one platform for generation, editing, and asset management, functioning as an “agent-like environment.”

Ownership & Usage Rights:

Manis explicitly states that users retain ownership of their creations and can use them for personal or commercial purposes, addressing a crucial concern for professional adoption.

Performance & Iteration Speed:

Image generation typically takes 10-30 seconds, but the primary benefit of Design View is increased iteration speed. The ability to fix details without full regeneration saves significant time and effort.

Conclusion:

Both GLM 4.7 and Manis Design View represent significant advancements in their respective fields. GLM 4.7 establishes a new benchmark for open-source coding agents, offering a stable and cost-effective solution for complex workflows. Manis Design View fundamentally alters the AI image editing experience, moving beyond the limitations of regeneration and enabling a more controlled and iterative creative process. These updates highlight a broader trend towards more practical, controllable, and integrated AI tools.