返回
原文
elvis@omarsar0
40AI 编辑部评分，满分 100
2026-08-01 05:50· 16分钟前
跳到正文
AI 摘要

DAIR.AI 的 Elvis Saravia 指出，模型 token 效率趋势始于 Qwen-3.7、Nemotron 和 DeepSeek-V4 preview，无需再绑定单一框架。

People haven't been paying enough attention.

No point in staying loyal to one harness anymore.

Btw， this token efficiency trend started with models like Qwen-3.7， Nemotron， and DeepSeek-V4 preview.

I've been testing， and I can confirm that lightweight harnesses like Pi and Hermes work better with frontier open models like Kimi K3.

ComposioWe ran Kimi K3 through 3 more agent harnesses (Pi Agent, OpenCode, and Codex) bringing the comparison to 6 harnesses across 26 agentic tasks. 2 things stood out...
智能体评测/基准
在 X 查看原推导出 Markdown
返回精选列表
40导出 Markdown
elvis · @omarsar0 · X·2026-08-01 05:50·16分钟前
在 X 看原推· x.com
AI 摘要

DAIR.AI 的 Elvis Saravia 指出，模型 token 效率趋势始于 Qwen-3.7、Nemotron 和 DeepSeek-V4 preview，无需再绑定单一框架。

People haven't been paying enough attention.

No point in staying loyal to one harness anymore.

Btw， this token efficiency trend started with models like Qwen-3.7， Nemotron， and DeepSeek-V4 preview.

I've been testing， and I can confirm that lightweight harnesses like Pi and Hermes work better with frontier open models like Kimi K3.

ComposioWe ran Kimi K3 through 3 more agent harnesses (Pi Agent, OpenCode, and Codex) bringing the comparison to 6 harnesses across 26 agentic tasks. 2 things stood out...
智能体评测/基准
在 X 查看原推x.com