AI Coding · Open Source · Frontend

Kimi K3.1 Pro in the Trenches: 9 Minutes to Fix a Stubborn Bug That Beat Other Models, Backed by 270K Lines of Code

By 孟健AI编程 · Oct 26, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

For an independent developer running a one-person shop, the practical question isn't benchmark scores but whether a model can debug a real, environment-specific bug end to end without hand-holding. K3.1 Pro's 9-minute zero-touch fix of an X editor cursor issue shows that frontier Chinese models are now delivering turnkey engineering work, not just code snippets — but its slow reasoning and aggressive quota consumption are real costs to budget for.

Summary

Moonshot AI's Kimi K3.1 Pro now ranks fourth globally on the Artificial Analysis composite leaderboard with 55 points, just behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and the not-yet-API'd Muse Wonder Full (56). On SWE Marathon v1.1 it tops the chart at 51.7%, beating Claude Opus 5.5's 46.3% and GPT-6 Astra's 42.5%. Terminal-Bench 4.0 shows 65.7%, a jump from K3's 13%, landing within 0.7 points of Claude Opus 5.5. API pricing drops about a third from the previous generation: domestic standard input at 13 yuan per million tokens and output at 68 yuan, with a high-speed tier costing twice as much.

Takeaways

— Kimi K3.1 Pro uses a mixture-of-experts architecture with 2.8 trillion total parameters, activating about 104 billion per inference, with a 1-million-token context window.

— Official commitment to open-source full model weights before November 4, 2026.

— AA composite score of 55 places it fourth globally; domestic rivals like DeepSeek and GLM sit around 45.

— SWE Marathon v1.1 score of 51.7% ranks first, above Claude Opus 5.5 (46.3%) and GPT-6 Astra (42.5%).

— Terminal-Bench 4.0 score of 65.7% improves from K3's 13% and is only 0.7 points below Claude Opus 5.5.

— API pricing is roughly one-third cheaper than previous generation: domestic standard input 13 yuan/M tokens, output 68 yuan/M; high-speed version doubles to 26 yuan input and 136 yuan output. USD: $2/$10 standard, $4/$20 high-speed.

— A persistent bug in the author's content distribution tool 'Ziliu' — 11 images always inserted as 10 with leftover garbage text — was fixed by K3.1 Pro in about 9 minutes using Kimi Code with --yolo mode.

— The fix approach: reproduce first, then change. K3.1 Pro ran the reproduction command, identified that the program was selecting a heading instead of the placeholder, and switched to simulating a mouse click so X's editor correctly sensed the cursor position.

— The model also built a landing page for the desktop version, using design skills 'taste' and 'impeccable', self-checking via browser screenshots at desktop and mobile sizes, catching and fixing 3 issues on its own.

— Known shortcomings: lags on ultra-long difficult tasks (safety pass@3 drops to 81.0% without offensive security data), slower reasoning (took over half an hour with one connection break for the landing page), and subscription quota drains quickly — Moonshot AI even issued an apology on Oct 26 for rate limits and unnecessary quota consumption.

Conclusions

The 9-minute bug fix is notable because the issue only reproduced in a real logged-in X page, which automated tests couldn't catch. The model had to read code, run a real browser, and verify the fix end to end — a genuinely different challenge from generating code in a vacuum.

The author's strict prompt structure (phenomenon, reproduction command, file scope, red lines) suggests that AI's reliability on messy engineering tasks hinges as much on how the task is framed as on model capability. The model didn't improvise; it followed a disciplined protocol.

The gap between K3.1 Pro and Claude Opus 5.5 on the AA composite score is small, but the speed and quota issues mean 'benchmark leadership' doesn't translate to smooth daily use. A model that's 10 points ahead can still feel frustrating if it burns your monthly subscription in one session.

Moonshot AI's explicit commitment to open-source the weights by 2026 is a strategic move that could shift how Chinese developers and enterprises adopt the model, potentially reducing vendor lock-in anxieties.

Concepts & terms

SWE Marathon v1.1

A benchmark that evaluates an AI model's ability to perform long, continuous coding tasks over many steps, simulating a real software engineering session rather than a single-shot code completion.

Terminal-Bench 4.0

A benchmark that tests an AI's performance on real terminal-based tasks with restricted environments, measuring how well it can handle command-line operations, file edits, and complex system interactions.

--yolo mode

A flag in Kimi Code that grants the AI agent permission to read/write files and execute commands without pausing for user confirmation at each step. It's meant for safe environments where changes can be rolled back via git.

pass@3

A security evaluation metric that measures the probability that at least one of three generated attempts passes a given safety test. A higher pass@3 indicates better adherence to safety constraints.

From the discussion

Featured comments

稳如狗

Your approach is to let Kimi fix bugs and build pages while running multiple rounds of testing and screenshots. Have you considered letting a strong model fix bugs and letting a cheap model handle the testing?

孟健AI编程

Tried that division of labor, but frontend testing relies on visual multimodality and long-context localization; cheap models often miss things and end up leaving noise in the context. As it stands, giving the whole troubleshooting-to-verification loop to the same strong model results in lower overall time cost.

See top comments, translated →

Source: juejin.cn ↗ Google Translate ↗ Backup ↗