MiMo v2.5
A lower-table result with a few useful bright spots: 41/151 tasks solved at least once, 17/151 solved in all three attempts, with the clearest wins around large Python/Django application repairs plus localized Go security-scanner changes. 这是一个排名靠后但仍有局部亮点的结果：151 题中至少一次解出 41 题，三次都解出 17 题；强项主要落在大型 Python/Django 应用修复以及边界相对清楚的 Go 漏洞扫描器改动。
How to read this result可以这样读
- MiMo v2.5 is best read as volatile explorer: rank #20, 41 reached tasks, 17 stable solves.MiMo v2.5 更适合读成探索型但波动较大：排名 #20，触达 41 题，稳定解出 17 题。
- Best suite signal: Open Library · release 013 at 7/10 (70.0%).最强 suite 信号：Open Library · release 013，7/10（70.0%）。
- Weakest visible area: qutebrowser · release 018 at 0/9 (0.0%).最弱可见区域：qutebrowser 浏览器 · release 018，0/9（0.0%）。
- Claude Code gives this row a different orchestration profile from OpenCode and Qoder, which is useful when comparing the same model family across shells.Claude Code 让这一行有别于 OpenCode 和 Qoder 的编排形态，适合观察同类模型跨 shell 的差异。
MiMo v2.5 is broad but volatile. It can touch 41/151 tasks, yet only 17 become 3/3 solves, so much of its value comes from retrying the same benchmark surface.
The closest family reference is MiMo v2.5 pro at rank #12. Compared with that row, this one is 3.65 points behind, with 3 fewer reached tasks and 8 fewer stable solves.
The suite split is asymmetric: Open Library · release 013 at 7/10 (70.0%) supplies the main body of wins, vuls · release 012 at 4/4 (100.0%) supplies the clean spike, and qutebrowser · release 018 at 0/9 (0.0%) is where that pattern stops. Claude Code gives this row a different orchestration profile from OpenCode and Qoder, which is useful when comparing the same model family across shells.
Read the bars as a volatility chart: Open Library · release 013 at 7/10 (70.0%) shows the upside, while the Pass^3 gap explains why the same row can feel much weaker on a single rerun.
The useful contrast is between Scan results miss Package URL (PURL) information in library output (future-architect/vuls · solved 3/3) and Ensure constant-like configuration values are immutable where consumed by autocomplete and related logic (internetarchive/openlibrary · solved 2/3). The model reaches both kinds of problems, but only one becomes dependable.
The audit trims 17 solved attempts from MiMo v2.5 but still keeps 80% of the solved set, so the suite shape remains useful even where individual wins are debatable.
The available audit keeps 69 of 86 initial solved attempts. Read this as a robustness check, especially when the audit sample is smaller than 453 attempts.当前可用复核保留了 86 次初始成功中的 69 次。这更适合作为稳健性检查，特别是在复核样本小于 453 次尝试时。
Use it when breadth matters more than deterministic replay. It can find openings around Open Library · release 013 at 7/10 (70.0%), but the 24-task reach gap says a second or third run may tell a different story. The 86/453 attempt score is best read as exploration bandwidth: 41 tasks are reachable, but many need retry luck.
Supporting suite table
| Suite | Repo | Solved | Pass^3 | Rate |
|---|---|---|---|---|
release-zh-012-future-architect-vuls |
future-architect/vuls | 4/4 | 1 | 100.0% |
release-zh-013-internetarchive-openlibrary |
internetarchive/openlibrary | 7/10 | 3 | 70.0% |
release-zh-003-ansible-ansible |
ansible/ansible | 4/10 | 0 | 40.0% |
release-zh-010-future-architect-vuls |
future-architect/vuls | 4/10 | 3 | 40.0% |
release-zh-016-internetarchive-openlibrary |
internetarchive/openlibrary | 2/5 | 0 | 40.0% |
release-zh-004-ansible-ansible |
ansible/ansible | 1/3 | 1 | 33.3% |
release-zh-018-qutebrowser-qutebrowser |
qutebrowser/qutebrowser | 0/9 | 0 | 0.0% |
release-zh-017-navidrome-navidrome |
navidrome/navidrome | 0/5 | 0 | 0.0% |
release-zh-005-flipt-io-flipt |
flipt-io/flipt | 1/10 | 0 | 10.0% |
release-zh-006-flipt-io-flipt |
flipt-io/flipt | 1/10 | 1 | 10.0% |
MiMo v2.5 的特点是覆盖不窄但波动较大。它能至少一次摸到 41/151 题，但只有 17 题能做到 3/3，因此很大一部分价值来自重试。
最接近的同系参照是排名 #12 的 MiMo v2.5 pro。和它相比，这一行最终分低 3.65 分，触达题少 3 个，稳定题少 8 个。
suite 分布是不对称的：Open Library · release 013，7/10（70.0%）贡献主要胜利，vuls 漏洞扫描器 · release 012，4/4（100.0%）贡献最干净高点，而qutebrowser 浏览器 · release 018，0/9（0.0%）标出这种模式停止的地方。Claude Code 让这一行有别于 OpenCode 和 Qoder 的编排形态，适合观察同类模型跨 shell 的差异。
这张图更像波动率图：Open Library · release 013，7/10（70.0%）展示上限，而 Pass^3 落差解释了为什么单次重跑会显得弱很多。
最有用的对比是 Library 输出中的扫描结果缺少 Package URL (PURL) 信息（future-architect/vuls · 3 次中成功 3 次）和 确保 autocomplete 和相关逻辑所消费的 constant-like configuration values 是 immutable 的（internetarchive/openlibrary · 3 次中成功 2 次）：模型都能触达，但只有前者变成可靠结果。
复核从 MiMo v2.5 中剔除了 17 次成功，但仍保留 80% 的成功集合，因此即便个别胜利有争议，suite 形状仍然有参考价值。
The available audit keeps 69 of 86 initial solved attempts. Read this as a robustness check, especially when the audit sample is smaller than 453 attempts.当前可用复核保留了 86 次初始成功中的 69 次。这更适合作为稳健性检查，特别是在复核样本小于 453 次尝试时。
当你更看重覆盖面而不是确定复现时，它更合适。它能在Open Library · release 013，7/10（70.0%）附近找到入口，但 24 题的覆盖-稳定差说明第二、第三次运行可能给出不同结果。86/453 的单次尝试成功数更像探索带宽：41 道题能触达，但很多仍需要重试运气。
支撑这个判断的 suite 表
| Suite | Repo | 解出 | Pass^3 | 通过率 |
|---|---|---|---|---|
release-zh-012-future-architect-vuls |
future-architect/vuls | 4/4 | 1 | 100.0% |
release-zh-013-internetarchive-openlibrary |
internetarchive/openlibrary | 7/10 | 3 | 70.0% |
release-zh-003-ansible-ansible |
ansible/ansible | 4/10 | 0 | 40.0% |
release-zh-010-future-architect-vuls |
future-architect/vuls | 4/10 | 3 | 40.0% |
release-zh-016-internetarchive-openlibrary |
internetarchive/openlibrary | 2/5 | 0 | 40.0% |
release-zh-004-ansible-ansible |
ansible/ansible | 1/3 | 1 | 33.3% |
release-zh-018-qutebrowser-qutebrowser |
qutebrowser/qutebrowser | 0/9 | 0 | 0.0% |
release-zh-017-navidrome-navidrome |
navidrome/navidrome | 0/5 | 0 | 0.0% |
release-zh-005-flipt-io-flipt |
flipt-io/flipt | 1/10 | 0 | 10.0% |
release-zh-006-flipt-io-flipt |
flipt-io/flipt | 1/10 | 1 | 10.0% |