AI 旗舰模型，还是你能选的日常使用模型吗？

前几天我写过一个关于 RRLabBench 的实测——给 9 个模型完全相同的编程任务，对比它们的效率和成本。具体细节翻上一篇就能看到。

A few days ago I wrote about RRLabBench — giving 9 models identical programming tasks and comparing their efficiency and cost. Flip back one post for the details.

对实测结果中 K3 的表现很疑惑。K3 的效率测试里排倒数第一：无论是解决同一个问题使用的轮次、消耗的时间都不排在前。而最终完成任务的成本，是同样完成任务的 Deepseek V4 Flash 的 16 倍。
外界对 Kimi K3 评价都很强，且很多评价来自 AI 研究领域。那是我忽略了什么？
在 K2.6 时代，我对 Kimi 的印象很好。当时 Kimi 发表过一篇"跨层注意力机制"的论文——通过这个机制，模型可以跨深度地选择性检索前面各层的信息。说白了，模型的不同层之间可以直接"对话"！这个机制让 K2.6 的深度思考模式有着非常高的质量。我的很多模型基础规则，就是靠它的深度思考模式帮我推理完成的。

K3's performance puzzled me. It ranked last in efficiency: neither rounds nor time favored it. And the cost to complete the same task was 16× that of DeepSeek V4 Flash.
The industry raves about K3, with much praise coming from AI research circles. What was I missing?
Back in the K2.6 era, I had a great impression of Kimi. They published a "cross-layer attention mechanism" paper — the model could selectively retrieve information across depth from earlier layers. In plain terms: different layers could directly "talk" to each other! This mechanism gave K2.6's deep-thinking mode exceptionally high quality. Many of my foundational model rules were reasoned out with its help.

到了 K3，一个更新的旗舰，为什么在日常 coding 任务上反而表现这么"差"？

So with K3, a newer flagship, why does it perform so "poorly" on everyday coding tasks?

花了些时间，翻了 K3 的技术架构和训练目标，答案渐渐浮出来了。不是 K3 变差了，是它和 K2.6 走的是两条不同的路。Kimi K3 有一个隐藏设定，很多人没注意到：它是全程 thinking 模式。发布时 reasoning_effort 只支持 max，没有浅层推理的开关。这意味着你让它改一行代码，它也会进入完整的推理流程：思考、验证、再思考。每一轮都深度推理，10 轮累积下来，时间就是同等模型的几倍。所以我在 RRLabBench 实测中 K3 的"124 秒"不是某几轮拖了后腿，是每一轮都匀速地"慢"。

After digging into K3's architecture and training objectives, the answer emerged. K3 didn't get worse — it took a fundamentally different path from K2.6. K3 has a hidden setting many overlooked: it's in full-time thinking mode. At launch, reasoning_effort only supported "max" — no shallow reasoning toggle. This means even for a one-line code change, it enters the full reasoning pipeline: think, verify, re-think. Every round runs deep inference. Stack 10 rounds and the time becomes several multiples of equivalent models. That "124 seconds" in my RRLabBench test wasn't a few outlier rounds dragging things down — every round was consistently "slow."

K3 垫底，不是因为它差了，是因为 RRLabBench 测的不是它被设计要干的事。而这种现象不是 Kimi 独有的，是整个 AI 大模型行业正在发生的事情。

K3 ranked last not because it's worse, but because RRLabBench wasn't testing what K3 was designed to do. And this phenomenon isn't unique to Kimi — it's happening across the entire large-model industry.

K3 的设计目标不是"和人讨论问题"，而是"独自完成长任务"。它的训练重心是长程 Agent 执行，连续几百步操作不能丢失上下文、不跑偏、不在某一步忘了前面发生的事。全程 thinking 在这种场景下不是浪费，是刚需：每一步都需要深度理解当前位置在整个任务链中的意义。目标变了，任务性质也变了。K3 能让一个模型运行几个小时，完成原本一个工程师一两天的工作量。它的目标客户是"无人值守的 CI 流程、金融合规审查、医疗报告生成"这类场景——在这些场景中，任务失败的代价至少是 token 费用的几百倍。

K3 wasn't designed to "discuss problems with humans" — it was built to "complete long tasks autonomously." Its training focused on long-range agent execution: hundreds of consecutive operations without losing context, drifting off-track, or forgetting earlier steps. Full-time thinking in this scenario isn't waste — it's a necessity: every step requires deep understanding of where you are in the overall task chain. Different goal, different task nature. K3 can run for hours, completing what would take an engineer one to two days. Its target customers are "unattended CI pipelines, financial compliance reviews, medical report generation" — scenarios where the cost of task failure is hundreds of times the token cost.

个人在日常工作处理、coding工作中还需要追旗舰吗？
日常任务的难度在过去两年并没有跟着旗舰能力一起上涨。两年前遇到的bug和今天遇到的bug，本质复杂度的提升微乎其微。而模型能力曲线在陡升，中档定价的模型 SWE-bench 都在 80% 以上。而这个分数在一年前是旗舰专属。
结论是：日常任务用旗舰模型就可能是浪费。
行业里的工程实践其实已经开始转变，现在流行的做法是先在便宜模型上测质量底线，只对实测失败的任务才升级到贵的。有些产品干脆把"日常走便宜模型、硬任务才升级"做成了基础设施。它背后的假设很简单：你不需要一个什么事都能做的模型，你需要一个知道什么时候该叫谁的系统。
K3 垫底这件事，不是 K3 的问题，也不是 RRLabBench 的问题。是我理解模型的方式需要更新了！以前我觉得"深度思考"就是好模型，不管什么场景开它就对了。现在我发现，深度思考是一种能力配置，不是一种质量标签。同样的技术，K2.6 用它来帮你推理规则，K3 用它来帮你跑长任务。定位不同，就不该用同一把尺子去量。

Do individuals still need flagship models for daily and coding tasks?
The difficulty of daily tasks hasn't risen alongside flagship capabilities over the past two years. Bugs from two years ago and bugs today — the fundamental complexity increase is negligible. Meanwhile, the model capability curve is shooting upward: mid-tier models now hit 80%+ on SWE-bench, a score that was flagship-exclusive a year ago.
The conclusion: using a flagship for daily tasks may be wasteful.
Industry practice is already shifting. The prevailing approach now: test quality baseline on cheap models first, only escalate to expensive ones when a task actually fails. Some products have turned "cheap model for routine, escalate for hard tasks" into infrastructure. The underlying assumption is simple: you don't need a model that does everything — you need a system that knows when to call whom.
K3 ranking last isn't a K3 problem, nor an RRLabBench problem. It's that my way of understanding models needed updating! I used to think "deep thinking = good model" and turned it on regardless of scenario. Now I see deep thinking as a capability configuration, not a quality label. Same technology: K2.6 used it to help reason out rules; K3 uses it to run long tasks. Different positioning — shouldn't be measured with the same yardstick.

所以问题变成：如何使用多个模型建立你自己的工作流，并能在任务不同阶段使用不同的模型的能力才是接下来我们需要考虑的问题！

So the question becomes: how to build your own workflow using multiple models, deploying different model capabilities at different task stages — that's what we need to think about next!

最后说一句：Kimi K3现在没有浅层推理的开关的问题，可能在未来更新中会做调整。因为Kimi K3的发展目标是做动态推理层级，就是说在完成任务时不是一直暴力深度思考。而是按需调整思考深度。而现在的一律暴力深度思考的方式估计是发布压力的原因…希望K3的动态推理更新后，它的表现会有所改善！

One last note: K3's lack of a shallow reasoning toggle will likely be addressed in future updates. K3's development roadmap targets dynamic reasoning levels — meaning it won't brute-force deep thinking throughout a task, but will adjust depth on demand. The current brute-force approach was probably a launch-pressure compromise... Hoping K3's dynamic reasoning update brings improved performance!