China’s New Self Improving Open AI Beats OpenAI

Key Concepts

• Agentic Workflows: AI systems designed to autonomously coordinate tasks, use tools, and manage multi-step processes without constant human intervention.
• Self-Evolving Models: AI architectures capable of autonomous performance tuning, error correction, and workflow optimization.
• Mixture of Experts (MoE): An architecture where only specific sub-networks (experts) are activated for a given task, increasing efficiency.
• Thought Compression: A technique where models learn to solve complex problems using fewer tokens while maintaining high performance.
• Parallel Reasoning: Running multiple agentic chains simultaneously to refine outputs and reduce latency compared to single, long-chain reasoning.
• Natively Multimodal: Models built from the ground up to process text, images, and UI elements simultaneously rather than through modular "bolted-on" components.

---

1. MiniMax: M2.7 and Self-Evolution

MiniMax has open-sourced M2.7, a high-performance model utilizing a Mixture of Experts (MoE) architecture.

• Performance: It excels in software engineering, scoring 56.22% on SWE-Pro and 57.0% on Terminal Bench 2. It demonstrates "system-level" understanding, capable of debugging production environments and identifying complex issues like missing index migrations.
• Self-Evolution: The model utilized an internal scaffold to perform over 100 autonomous rounds of self-improvement. It tuned hyperparameters (temperature/penalty) and implemented loop detection, resulting in a 30% performance boost.
• Real-World Application: MiniMax currently uses M2.7 to handle 30–50% of its own reinforcement learning team workflows. It also performs junior-analyst tasks, such as synthesizing annual reports and building revenue models.

2. Runnable: Run Claw

Runnable introduced Run Claw, an AI agent integrated directly into communication platforms (Slack, Telegram, Discord).

• Shift to Delegation: Unlike traditional "prompt-box" AI, Run Claw is designed to ask clarifying questions, build execution plans, and perform tasks in the background.
• Ecosystem Integration: Runnable functions as an all-in-one platform capable of generating websites, managing databases, handling Stripe payments, and providing AI voice support. The company reports $2 million in ARR, highlighting the market shift toward integrated, agentic toolsets.

3. Google: Mixboard Evolution

Google is transforming Mixboard from an experimental image canvas into a collaborative workspace.

• Voice Control: The platform is integrating full voice-command capabilities, allowing users to generate, rearrange, and swap content without manual interface interaction.
• Workflow Integration: A key feature is the ability to export board sessions into structured documents, bridging the gap between brainstorming and formal documentation.

4. OpenAI: Unified Codex Agent App

OpenAI is developing a unified platform that merges ChatGPT, the Atlas browser, and coding tools.

• Scratchpad & Heartbeat: The app features a "scratchpad" for parallel task execution and a "heartbeat" system, which maintains persistent connections for long-running, background agentic processes.
• Strategic Direction: This move aims to reduce friction by centralizing chat, coding, and browsing into a single agent-managed environment.

5. Meta: Muse Spark

Meta’s Muse Spark is a natively multimodal model developed by their Super Intelligence Labs.

• Technical Efficiency: It is 10x more compute-efficient than Llama 4 Maverick.
• Reasoning Frameworks:
  • Thought Compression: Optimizing token usage for complex problem-solving.
  • Contemplating Mode: Running multiple agents in parallel to refine answers, which significantly boosts performance in scientific and health-related benchmarks.
• Benchmarks: It achieved a 42.8 score on Health Bench Hard (outperforming Opus 4.6 and Gemini 3.1) due to training data curated by over 1,000 physicians. However, it currently lags in abstract reasoning (ARC AGI 2).

---

Synthesis and Conclusion

The industry is rapidly transitioning from generative AI (creating content) to agentic AI (executing work). The common thread across these updates—from MiniMax’s self-evolving M2.7 to OpenAI’s persistent "heartbeat" agents and Meta’s parallel reasoning—is the move toward autonomous, long-running workflows. These systems are no longer just answering questions; they are acting as systems engineers, analysts, and project managers that can operate across multiple files, tools, and platforms with minimal human oversight. The competitive advantage is shifting toward platforms that can reduce "context switching" and provide end-to-end task completion.