Introducing SWE-2: RL at multi-trillion-parameter scale, 50% FrontierCode at 64% less cost

ID: 69669dae
原文链接: https://cognition.com/blog/swe-2
作者: Cognition
发布日期: 2026-09-10
条目分类: coding
来源类型: article
标签: swe-2, cognition, reinforcement-learning, coding-model, field-note
质量评分: 4/5
简评作者: openclaw
抓取时间: 2026-09-11 08:17 (UTC+8)

中文导读

Cognition 发布 SWE-2，称其为最强 coding 模型：FrontierCode 1.1 Main 拿 50.0%，距 Fable 5.1（50.9%）一分之内而便宜 64%；距 GPT-6 Astra（53.3%）只差几分而成本只要四分之一；DeepSWE 1.1 拿 73.0%（仅次于 Astra 74.1%），Terminal-Bench 2.1 拿 92.8% 领先对照表里所有模型。

训练侧两个第一次：

1. 首次把 RL 扩到万亿参数级：后训练基座是 Kimi K3（2.8T 参数，已经为 agentic coding 做过大量 RL）。SWE-2 在这个基础上再加 5–6 分，并把 K3 的整条成本-性能前沿整体移动。 2. 单次运行训练所有 reasoning-effort 档位：新 RL 算法给每个档位施加线性成本惩罚（cost penalties），让一次训练同时推进所有档位的成本-性能前沿，而不是每个档位单独训练。

基准对照（官方表格，FrontierCode 1.1 Main / DeepSWE 1.1 / Terminal-Bench 2.1 / Terminal-Bench 4）：SWE-2 = 50.0 / 73.0 / 92.8 / 27.3；Kimi K3 = 44.2 / 68.5 / 88.3 / 21.5；Grok 4.6 = 48.0 / 67.5 / 88.4 / 20.3；Fable 5.1 = 50.9 / 67.4 / 91.4 / 55.8；GPT-5.6 Sol = 47.5 / 72.7 / 88.8 / 37.3；GPT-6 Astra = 53.3 / 74.1 / 89.9 / 57.9；SWE-1.7 = 42.0 / 37.7 / 81.5 / 7.6。

值得注意的短板：Terminal-Bench 4（长程任务）SWE-2 只有 27.3%，明显落后 Astra 57.9% 与 Fable 5.1 55.8%——"frontier 差几分、便宜一个量级"的性价比曲线在短程任务上成立，长程任务上代差仍在。

Why it matters

RL 训练规模首次进入 multi-trillion 参数区间，且"单次训练全部 effort 档位+线性成本惩罚"是可借鉴的 recipe 级信息。
基座选已充分 RL 过的 Kimi K3 仍能加 5–6 分——RL 在 agentic coding 上的 headroom 还没打满，这对所有做 coding agent 的团队是直接信号。
价格信号：以 Fable 5.1 的 36% 价格拿到其 98% 的分数，agentic coding 的成本曲线进入快速下移通道。

要点摘录（opencli web 抓取，2026-09-11）：

SWE-2 post-trained from Kimi K3 (arXiv 2607.24653), a 2.8T-parameter model
FrontierCode 1.1 Main: score = weighted aggregate of rubric items; blocking criteria fail = 0; cost = mean USD per rollout
Leaderboard: cognition.com/frontiercode

原文摘录

Today we're introducing SWE-2, our most advanced coding model yet. It pushes the Pareto frontier of capability and cost, achieving 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while being 64% cheaper.

With SWE-2, we scaled RL to the multi-trillion-parameter regime for the first time, building on the SWE-1.7 training infrastructure and recipe. The key addition is an RL algorithm that trains all reasoning-effort levels in a single run, advancing the whole cost-performance frontier.

关键信息

文章标题：Introducing SWE
发布方：Cognition
发布时间：2026-09-10
原文：https://cognition.com/blog/swe-2
关联标签：swe-2, cognition, reinforcement-learning, coding-model

English Summary

Cognition introduces SWE-2, its most advanced coding model: 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 (50.9%) while 64% cheaper, a few points from GPT-6 Astra (53.3%) at a quarter of the cost, and 73.0% on DeepSWE 1.1. For the first time RL was scaled to the multi-trillion-parameter regime, post-training Kimi K3 (2.8T parameters, already extensively RL-trained for agentic coding) with a new RL algorithm that trains all reasoning-effort levels in a single run using per-level linear cost penalties, advancing the whole cost-performance frontier (adding 5-6 points on many benchmarks over K3). Terminal-Bench 2.1: 92.8% (best in its table); Terminal-Bench 4: 27.3%, still far behind Astra 57.9% and Fable 5.1 55.8%.

Obsidian Notes

正文由 opencli web read 抓取全文（2026-09-11，UTC+8），导读、数字与摘录均锚定在抓取内容上。
与条目 ba08af0c（DeepSeek V4.1 Flash 降价）同周发布，coding 模型价格竞争可对照阅读。