记录模型数据、评测流程和网站展示方式的持续演进。Tracking the evolution of model data, benchmark workflows, and the way results are presented.
新增 83.3 GB 的 IQ1_M 极低比特量化模型完整结果。能力上限 94.4、实用得分 91.4，ToolCall 与 BugFind 各 100 分，HermesAgent 86 分；50 题中最终通过 45 题，仅需 3 次重试。Added complete results for the 83.3 GB IQ1_M ultra-low-bit quant model. Max score 94.4, effective 91.4; ToolCall and BugFind both reach 100, HermesAgent 86; 45 of 50 scenarios ultimately pass with only 3 retries.
重构数据采集、评分、页面生成和发布前校验流程。现在以规范化快照作为唯一数据源，自动合并同一模型的思考与无思考结果，并用能力上限与实用得分同时呈现模型表现。Rebuilt the data collection, scoring, page generation, and pre-release validation workflow. Normalized snapshots are now the single source of truth, thinking and no-thinking results are merged by model, and both max and effective scores are presented.
首页、排行榜、模型详情页和更新日志完成全新视觉升级；模型 LOGO、字体与方法插图全部本地化，旧详情页地址继续通过兼容跳转访问。The home page, leaderboard, model details, and changelog received a complete visual refresh. Model logos, fonts, and method artwork are local, while legacy detail URLs remain available through compatibility redirects.
新增思考与无思考完整结果。思考版能力上限 93.7、实用得分 87.7；无思考版能力上限 91.0、实用得分 88.0。Added complete thinking and no-thinking results. Thinking reaches 93.7 max and 87.7 effective; no-thinking reaches 91.0 max and 88.0 effective.
思考和无思考结果合并为同一个模型入口，并保留两套完整的 50 题明细。Merged thinking and no-thinking results under one model entry while preserving both complete 50-scenario tables.
查看模型详情View model details →修复大分数字段、正负差值以及完整题目表格的显示问题。Fixed headline scores, signed deltas, and complete result-table rendering.
明确结果用于参考和简单对比，并补充速度数据与硬件环境的适用边界。Clarified that results support reference and rough comparison, with scope notes for speed and hardware context.
加入系列模型的多个变体，并统一它们在首页与详情页中的命名和来源信息。Added multiple series variants with consistent naming and source information across home and detail pages.
开始集中记录模型测试、数据调整和网站功能迭代。Started tracking model benchmarks, data adjustments, and site improvements in one place.