Smlay_ero

笔记

Notes

发在 X 上的长文。全文在原帖里读，这里只放标题和摘要。

Long posts published on X. Read the full piece there; this page keeps the title and a short excerpt.

2026-09-03 · X

Beyond Q&A: Stress-Testing a Finance-Tuned Model on Three Real Investment Workflows

金融工作面对的信息一直在变，定义也很少对齐。一次分析往往同时跨宏观、行业和公司。这篇把一个金融微调模型放到三条真实投研工作流里压测，而不是停在问答演示。

Financial work deals with information that shifts constantly and definitions that rarely line up. A single analysis often spans macro, sector, and company. This piece stress-tests a finance-tuned model on three real investment workflows, instead of stopping at Q&A.

在 X 阅读Read on X
2026-08-28 · X

A 1.3B-Active-Parameter Model Just Beat a 120B Giant. The Arms Race Is Over.

7.9B 总参数、每 token 只激活 1.3B 的模型，在公开指数上超过了 120B 的 gpt-oss。双卡 RTX 5090 + vLLM 本地实测了吞吐和稳定性。对企业来说，关键不是最大的模型，而是够用、跑得起、管得住。

A 7.9B model that activates 1.3B parameters per token scored above 120B gpt-oss on a public index. We ran it locally on dual RTX 5090s with vLLM. What enterprises need is not the biggest model — one that is good enough, cheap enough to run, and governable.

在 X 阅读Read on X