Bonsai Ternary Bonsai 2 27B基准表现仅保留Qwen3
bonsai's document reveal how much cherry picked their headlines are
bonsai claim 98.2% intelligent retained, but their own documents show Ternary Bonsai 2 27B reaches 52.8 and 60.8, respectively, compared with 69.7 and 80.6 for Qwen3.5-27B, retaining roughly three quarters of the full-precision performance on both benchmarks.
Bonsai 声称其智能保留率为 98.2%，但其自身文档显示，Ternary Bonsai 2 27B 在两项基准测试中分别达到 52.8 和 60.8，而 Qwen3.5-27B 分别为 69.7 和 80.8，在两项基准上均保留了全精度性能的大约四分之三。
that qwen3.5 is a typo cause these are qwen3.8 numbers, altho qwen3.5 numbers are
qwen3.5 是个笔误，因为这些是 qwen3.8 的数据，尽管 qwen3.5 的数据是
- bonsai_2/q3.5 52.8 / 41.6 = 126.9%
- bonsai_2/q3.5 60.8 / 72.4 = 84.0%
Long-context and coding performance. This release also delivers on the roadmap set out in our initial Bonsai 27B release [2], where we identified long-horizon, tool-driven software engineering as the next major capability to improve. With Ternary Bonsai 2 27B, that progress now shows up directly in agentic performance. Evaluated for the first time on Terminal-Bench 2.1 and SWE-bench Verified, the Ternary Bonsai 2 27B reaches 52.8 and 60.8, respectively, compared with 69.7 and 80.6 for Qwen3.5-27B, retaining roughly three quarters of the full-precision performance on both benchmarks.
长上下文与编码性能。此次发布也兑现了我们在初始 Bonsai 27B 发布 [2] 中制定的路线图，当时我们将长周期、工具驱动的软件工程确定为下一个需要提升的主要能力。借助 Ternary Bonsai 2 27B，这一进展现在直接体现在代理（agentic）性能上。在 Terminal-Bench 2.1 和 SWE-bench Verified 上首次进行评估，Ternary Bonsai 2 27B 分别达到 52.8 和 60.8，而 Qwen3.5-27B 分别为 69.7 和 80.6，在两项基准上均保留了全精度性能的大约四分之三。
link to their whitepaper on github, it is on page 7
他们在 GitHub 上的白皮书链接，在第 7 页
also 3.8 35b qwhen? plsss
还有 3.8 35b qwhen？求求了
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力