New data from the Harness Leaderboard has provided a comparative analysis of several prominent AI models, testing their capabilities against the Terminal Bench 2.1 and Deep SWE 1.1 benchmarks. The results include performance metrics for Kimi Code, Hermes, Pi, Claude Code, and Codex. Notably, the leaderboard identified mini-swe-agent as a top performer, a result that has drawn attention from observers on social media platforms like Reddit, where some users noted the model’s unexpected prominence.

The benchmarking process highlights the increasing complexity of evaluating “agentic” AI—systems designed to perform autonomous tasks. While some users on Reddit expressed disappointment regarding the absence of models such as DeepSeek or Cursor in these specific results, the data underscores a growing industry focus on specialized coding and software engineering benchmarks.

Acceleration in AI Research and Development

The pace of AI research and development (R&D) appears to be accelerating due to the integration of AI-assisted automation. Ryan Greenblatt of METR has estimated that AI R&D acceleration has reached levels previously projected for March 2027. According to Greenblatt’s analysis, AI-driven research and development (RSI) is currently speeding up overall AI progress by approximately 2x.

Greenblatt’s modeling suggests that the impact on software-specific R&D is even more pronounced. He estimates that RSI may be accelerating AI software progress by roughly 4.5x compared to scenarios without AI automation. This acceleration is attributed to the fact that a significant portion of AI progress is driven by software improvements rather than just compute scaling. However, Greenblatt noted that because certain portions of the underlying reports are redacted, some conclusions regarding the exact drivers of this speed-up rely on the rigor and trust placed in the authors.

Efficiency Gains in AI Memory Systems

In a separate development, Stanford researchers have introduced a new AI memory system designed to improve how models recall information from extensive conversation histories. The system utilizes only 16 memory tokens to manage data, which reportedly reduces document-token usage by up to 4,212×.

In a test involving 512 documents, the system ran 27× faster than existing methods and outperformed current approaches on five out of eight benchmarks. This development addresses the high computational costs associated with managing long-context windows in large language models.

Global Competition and Infrastructure Expansion

The landscape of AI development continues to see significant activity in China, despite ongoing U.S. restrictions on talent and chip technology. Reports indicate that Chinese AI models are maintaining high levels of activity, with some estimates suggesting Chinese models reached 4.69 trillion weekly tokens, surpassing U.S. totals.

Several Chinese firms are expanding their specialized AI capabilities:

  • Alibaba has released the open-source Qwen3 and Qwen3-Coder, the latter designed for complex coding workflows.
  • Kuaishou Technology has unveiled a video generation model intended to compete with OpenAI’s Sora.
  • 360 Security Technology is developing tools to match the capabilities of Anthropic’s Mythos.
  • Baidu reported the successful operation of a cluster comprising 30,000 self-developed P800 Kunlun chips.

Additionally, the integration of AI into physical and scientific sectors is expanding, with reports of AI-driven drug screening platforms and the use of quantum computers to fine-tune billion-parameter models.