Tech • AI • Robotics
China’s New Self Improving Open AI Beats OpenAI
6/10TL;DR
MiniMax released M2.7, a powerful self-optimizing open-source AI model for software engineering and office tasks; alongside, major AI updates emerged from Runnable, Google Mixboard, OpenAI, and Meta’s Muse Spark, highlighting rapid advances in AI agent capabilities and multimodal systems.
KEY POINTS
MiniMax Launches M2.7: A Strong Open-Source Model with Self-Evolution
MiniMax has fully open-sourced the weights for M2.7, its most capable model to date, hosted on Hugging Face. The model leverages a mixture of experts architecture, activating only necessary subcomponents during task execution to improve efficiency. It excels primarily in software engineering, office work, and managing “agent teams” — essentially coordinating multiple AI agents to collaboratively tackle complex assignments end-to-end. In software engineering benchmarks close to real-world scenarios, M2.7 scores competitively at 56.22% on SWE-Pro (near GPT-5.3 Codex level), 57.0% on Terminal Bench 2, and 39.8% on NL2 Repo, demonstrating a strong understanding of entire codebases rather than isolated snippets. Its broader engineering skill benchmarks were also impressive, including 76.5% on SWE-Multilingual and 52.7% on MultiSWE-Bench.
Production Debugging and Real-World Utility
Beyond coding, M2.7 stands out by assisting in live production debugging, identifying root causes for issues like missing database migration indexes, correlating monitoring spikes to deployment actions, and delivering fixes that can drastically reduce incident recovery to under 3 minutes. This situates it closer to a Site Reliability Engineer (SRE) role than typical code generators.
Self-Improvement through Autonomous Optimization
A pioneering aspect of M2.7 is its internal self-evolution framework, where it autonomously conducted over 100 optimization cycles analyzing its failures, refining scaffold code, tuning parameters (like temperature and penalties), and enhancing loop-detection to avoid redundant agent actions. This process yielded a 30% uplift in internal evaluation performance, marking a significant AI-driven improvement loop rarely seen publicly.
Handling Complex Workflows and Office Tasks
M2.7 supports extended, multi-agent workflows with a 97% skill compliance rate over tasks exceeding 2000 tokens and an overall accuracy of 62.7% on MiniMax’s MM-Claw evaluation. It also addresses professional office workloads, scoring an ELO of 1,495 on GDP-Val AA, ranking as the top open-source model for expert-level tasks, able to analyze financial documents, forecast revenues, and generate detailed presentations and reports akin to junior analysts.
Strong Performance on Machine Learning Competitions
In the MLE-Bench Light—a collection of 22 machine learning competitions executable on a single Nvidia A30 GPU—M2.7 demonstrated competitive prowess by earning 9 gold, 5 silver, and 1 bronze medals in a 24-hour run, averaging a 66.6% medal rate, closely rivaling models such as Gemini 3.1 and just behind GPT-5.4 and Opus 4.6.
Runnable Introduces Run Claw: Cloud AI Agent Integration in Team Chats
Runnable launched Run Claw, a cloud-based AI agent accessible via messaging platforms like Slack, Telegram, and Discord. This system moves beyond traditional prompt-response interfaces by querying users for clarification, crafting strategic plans, and autonomously executing tasks. Runnable’s platform already supports creating websites, videos, pitch decks, and more, integrating with tools such as Google, Slack, Notion, GitHub, and Shopify. Having surpassed $2 million Annual Recurring Revenue (ARR) with frequent feature rollouts, Runnable is positioning itself as a unified AI service enabling seamless task delegation directly through team communication channels.
Google Evolves Mixboard with Full Voice Control and Collaborative Features
Google announced significant enhancements to Mixboard, transforming it from an image-focused canvas into a hybrid collaborative workspace incorporating stickers, voice notes, geometric shapes, and markers. Most notably, they introduced full voice-control capabilities allowing users to generate images, rearrange elements, and operate the entire interface hands-free, built using the infrastructure underlying Google’s Stitch voice tool. Additionally, Mixboard supports experimental PDF exports to automatically convert brainstorming sessions into formal documents, streamlining the transition from ideation to documentation. These features might debut at the upcoming Google I/O conference.
OpenAI Developing a Unified Codex Agent Application
OpenAI is designing a comprehensive Codex agent app to merge ChatGPT, the Atlas browser, and coding capabilities into one platform. The app includes a "scratchpad" interface allowing multiple Codex tasks to run concurrently and harnesses managed agents that persistently operate in the background with periodic updates. Evidence of a “heartbeat” system indicates these agents maintain active long-running processes, resembling functionality seen in Runnable's Run Claw. This aligns with broader industry trends as Anthropic introduces a similar multi-agent system called Conway. Speculation surrounds a forthcoming model, possibly GPT-5.5 “Glacier,” hinted through snowflake emojis by OpenAI staff, which may coincide with this platform rollout.
Meta Releases Muse Spark: A Natively Multimodal AI from Scratch
Meta unveiled Muse Spark, developed by its Super Intelligence Labs as a multimodal AI model designed jointly for text and images rather than adding vision on top of a language-only model. In benchmark tests focusing on UI element recognition (Screen Spot Pro), Muse Spark achieves 72.2% to 84.1% accuracy, outperforming Opus 4.6 Max and GPT 5.46 High in base setups. Its training advances include revamped pre-training for tenfold better compute efficiency compared to Llama 4 Maverick, stable reinforcement learning improvements, and innovative “test-time reasoning” where multiple agents generate, refine, and merge answers in parallel (“contemplating mode”) to boost capability without long latency. This mode delivers top marks on complex tasks, scoring 58.4 on Humanity’s Last Exam (tool-augmented), narrowly trailing GPT 5.4 Pro.
Domain-Specific Strengths and Weaknesses of Muse Spark
Muse Spark demonstrates outstanding performance in health-related benchmarks (Health Bench Hard: 42.8) massively surpassing competitors, bolstered by data curated with input from over 1,000 physicians. In software engineering, it performs strongly but trails leaders with a 77.4% on SWE Bench Verified. However, in abstract reasoning tasks (ARC AGI 2), Muse Spark scores 42.5, considerably behind Gemini and GPT 5.4 models which score over 76, signaling ongoing work needed in this area.
The AI Landscape is Rapidly Evolving Towards Agent-Based, Collaborative, and Multimodal Systems
These updates collectively emphasize a shift from simple prompt-based AI generation to fully autonomous AI agents capable of multi-step workflows, complex coordination, self-optimization, and multimodal understanding. Models and tools are increasingly integrated into real-world workflows spanning code debugging, financial analysis, collaborative ideation, and continuous learning. This fast-paced evolution highlights growing competition among AI providers striving to deliver versatile, generalist agents embedded deeply into users' professional ecosystems.
The AI field continues to witness swift advancements with MiniMax’s M2.7, Runnable’s Run Claw, Google’s Mixboard, OpenAI’s unified Codex app, and Meta’s Muse Spark leading innovation in agent-based and multimodal AI technologies.
Explain this