Kimi K2.7 Code In-Depth Testing - Comprehensive Analysis of Code Generation Capabilities

Friends, Kimi has updated again.

This release is Kimi K2.7 CodeIt uses the same architecture as Kimi K2.6 and is a Coding-specific version that is trained on top of Kimi K2.6.

According to the official introduction, this upgrade mainly focuses on the following aspects:

By reducing overthinking, the usage of reasoning-tokens is reduced by 30% compared to the previous generation Kimi K2.6, meaning faster response and lower costs..

Benchmark tests also showed significant improvement.According to data released by The Dark Side of the Moon, Kimi K2.7 Code's performance in certain specific tasks has begun to approach... GPT-5.5 and Opus 4.8.

I've seen quite a few comparison reviews from friends these past couple of days. In the same...Prompt wordsBelow, some of the physical animations and front-end pages generated by Kimi K2.7 Code are even more expressive than... GPT-5.5 is better, and the actual cost is less than that. GPTOne-third of -5.5.

Since it's aimed at Coding and Agent A custom-tuned version tailored to specific scenarios, suitable for everyday programming tasks and...intelligentbodySwitching to this new model during the transition would undoubtedly be more appropriate. Furthermore, the official announcement states...A 6x high-speed mode will also be launched soon..

Let's test it out together.

This review uses Kimi Code + Kimi K2.7 Code, but it can also be used through the Kimi API.

SVG image

Let's try a lighter task first.

Prompt wordsPlease generate an SVG code for a cute panda holding a hamburger in both hands, as detailed as possible.

At first glance, the visual aesthetics are on point. The panda's expression, the hamburger filling, the sesame seeds, the paw pads... all these details are quite complete.

However, upon closer inspection, the panda's arms are not connected to its body, and the overall structural relationship still needs to be optimized.

SVG animation

Now that we've finished creating the static SVG, let's take a look at the animation effect.

Prompt words: Draw an animation of the tortoise and the hare race using SVG.

The graphics are decent and it runs normally, but the turtle and the hare are facing the wrong way, making it look like they're running backwards, and the hare ultimately wins, which doesn't quite fit the classic story.

Let's continue trying to fix the Kimi K2.7 Code:

Prompt wordsPlease fix two issues in the current Tortoise and Hare Race SVG animation:

1. The result of the match is incorrect.

* The current result is "The rabbit wins". Please change it to "The tortoise ultimately wins".

* It must conform to the classic tortoise and hare race logic: the hare is faster, but will stop to sleep or rest along the way; the tortoise is slower but keeps moving forward and is the first to reach the finish line.

* don't wantSimpleTo make the rabbit's speed slower than the tortoise's, the tortoise must overtake it by utilizing the rabbit's behavior of resting/sleeping midway. 2. Incorrect animal orientation.

* The current movement direction of the tortoise and the hare is correct; they are both running from left to right towards the finish line. Do not modify the movement path or the incrementing logic of the x-coordinate.

The problem is that the animals' heads are facing the wrong direction. Please only correct the characters' orientation so that both the turtle and the rabbit's heads are facing to the right, which is towards the finish line.

* If scaleX(-1), transform, animateMotion, path, or grouping are used

Please ensure that you only adjust the character's own visual orientation and do not reverse the direction of movement.

* Animals are prohibited from running backwards, running with their backs to the finish line, or moving to the right while their bodies are facing left.

After fixing it, please output the complete, runnable code.

This case study is perfect for testing whether the model truly understands the process logic within the animation. If the model only changes the result, it's easy to be lazy and simply slow down the rabbit's speed.

Kimi K2.7 Code breaks down the rabbit into three phases: starting ahead, sleeping and stagnating, and waking up and sprinting. The tortoise, on the other hand, maintains a steady pace and eventually overtakes the rabbit while it is sleeping.

This indicates that the model is beginning to shift from outcome-oriented optimization to process-level behavioral modeling, which can satisfy constraints by constructing state changes in time series, allowing the results to be generated naturally by the process, in accordance with the classic causal chain structure.

Prompt words[Topic] Volcanic Eruption

Objective: To create an interactive volcanic eruption scenario where clicking on the crater triggers an eruption.

Required features:

Click on the crater area to trigger a volcanic eruption.

Eruptions must include magma columns, Martian particles, volcanic ash, and lava flows; drawing only magma is not permitted.Simpleflame.

The sky and mountains were illuminated by a bright red-orange light at the moment of eruption, which then gradually faded.

A continuous system of volcanic ash particles requires the dust to rise and drift with the wind.

The bottom needs to include a silhouette of the terrain, such as a ridge, forest, village, or rocky wasteland.

It supports continuous clicking, and multiple bursts can be stacked on the same screen, with each burst decaying independently.

Must-do visuals:

The magma is mainly bright yellow/orange-red, gradually darkening at the edges.

The Martian particles need to exhibit parabolic motion and a gradual extinguishing effect.

The lava flow needs to slide slowly down the hillside, leaving a glowing trail.

Volcanic ash needs to have a sense of layering, it can't just be ordinary smoke dots.

Bonus points:

The magma ejection is simulated using a particle system, with random perturbations in velocity, direction, and magnitude.

The cracks on the mountain surface have a glowing effect; the stronger the eruption, the brighter the cracks.

Slight ground vibrations or camera shakes simulate an explosive impact.

In the distance, volcanic ash clouds slowly spread across the sky.

When lava comes into contact with the ground, it produces smoke and a cooling, darkening effect.

The eruption intensity increases with the frequency of eruptions, resulting in higher magma levels, more sparks, and thicker ash clouds.

After a period of inactivity, the volcano gradually cooled down and returned to calm.

Kimi K2.7 Code uses pure native Canvas to implement special effects, including lava, sparks, volcanic ash, lava flows, smoke, and distant gray clouds. Click frequency also affects eruption intensity, and strong light, cracks, and vibrations are also implemented. Multiple eruptions are not independent events, the lava flows are visually weak, and the logical loop is well-closed.

The downside is that the magma column is not well-formed, and the lava flow is relatively weak.

3D scene

Prompt words[Topic] Magic Forest Cabin

Objective: To build a web-based 3D isometric scene using HTML + JavaScript + Three.js: a little fox sitting at the entrance of a mushroom house, next to a glowing magic lamp.

Image requirements:

The scene is presented from an isometric perspective and has a clear 3D spatial hierarchy.

The little fox is adorable and sits next to the mushroom house.

The Mushroom House features a rounded roof, doors and windows, and fairytale-style details.

The magic lamp needs to be translucent and emit light, and become the visual focus of the image.

The ground contains grass, small flowers, stones, and a small number of firefly particles.

The lighting is soft, with ambient light, contour light, and localized illumination effects.

The materials should be differentiated: fox fur, mushroom house, glass lampshade, grass, and stone have different textures.

Technical requirements:

Outputs complete and executable HTML code.

Use Three.js to implement 3D effects on web pages.

It does not use Blender and does not rely on external image assets.

Use geometry, materials, lighting, and particles to build the scene as much as possible.

The code structure is clear, including scene initialization, camera, lighting, model creation, particle effects, and animation loop.

This type of task tests not only whether you can write Three.js, but also whether the model can handle scene initialization, geometry combination, materials, lighting, particles and animation loops at the same time.

Kimi K2.7 Code canfastThe complete scene structure is built, making it suitable for the first version of a 3D creative concept.

However, to truly create a high-precision work, camera composition, model proportions, lighting effects, and material details still require manual fine-tuning.

game

Prompt wordsCreate a Minecraft remake.

In less than five minutes, Kimi K2.7 Code recreated a simplified version of Minecraft. It not only completed terrain generation but also implemented first-person perspective control, as well as block destruction and placement interactions, essentially completing the core closed loop of a voxel sandbox.

UI Design

Prompt wordsCreate a modern Apple-style login page.

The layout, dark texture, fixed frosted glass navigation, large font titles, buttons, and scrolling effects all give it an Apple-like feel at first glance.

The overall completion level is close to the style reproduction level, but it is still not as Apple-style in terms of product credibility and information density control.

The device images on the first screen are just gradient cards, without any real devices or interfaces as visual anchors; while the glowing ball and SaaS cards make the screen richer, they weaken the emphasis on real product visuals and restrained expression in Apple's style.

Bug Detection and Repair

This is where a programmer's true skills are most tested. I'll use a real-life example...open sourceTest Kimi with the logic code for generating packaged filenames in the project's File Browser.

Prompt words:

Please verify the following true statement.open sourceThe project code underwent a security audit, and a deployable fix was provided.

background:

The project is a file browser. Its functionality involves packaging files into a zip/tar archive when a user downloads a folder. Please thoroughly review the filename generation logic within the archive for any security risks.

Require:

Determine if a vulnerability exists, and specify the vulnerability type and risk level.

Indicate the specific code location where the vulnerability occurred.

Explain the cause of the vulnerability and its potential impact on different operating systems.

The solution and key code patches are provided.

Provide a pre-deployment testing plan, including unit testing, security testing, cross-platform testing, and regression testing.

Provide production deployment solutions, including version upgrades, canary releases, log monitoring, user notifications, and rollback plans.

The answer must be executable; you can't just say "strengthen validation".

Code:

nameInArchive := strings.TrimPrefix(path, commonPath)

nameInArchive = strings.TrimPrefix(nameInArchive, string(filepath.Separator))

nameInArchive = filepath.ToSlash(nameInArchive)

Kimi's answer to K2.7 Code was generally acceptable, but not outstanding.

The identification of the vulnerability type was correct, and the risk point was also identified as a Zip Slip issue, which is related to the compressed file name.

Kimi K2.7 Code uses filepath.Rel and filepath.Clean to standardize paths, adds checks for backtracking paths (..), intercepts absolute paths, and unifies them into the / format used by zip/tar.

This indicates that the model has begun to use path semantics to handle security issues, demonstrating a certain level of engineering awareness.

It provides deployment solutions including testing, canary deployments, monitoring, and rollback. The engineering approach is quite comprehensive.

The main problem is that Kimi K2.7 Code does not establish a unified path standardization benchmark before making safety judgments on paths in multiple intermediate states, resulting in the mixing of original paths, Rel results, and Clean results, and inconsistent judgment semantics.

This fix is on the right track, but it fails to address the most critical issue of the real vulnerability: the "Windows backslash cross-platform path semantics".

landing page

Overall, the design has been able to capture the original design style at first glance. The key visual elements such as the purple color scheme, landscape illustration language, rounded corners of cards, and modular partitioning have been basically reproduced, and the style consistency is achieved.

The original purple has obvious warm and cool variations, with a gradual transition from light purple to dark purple. The replica version of purple is more uniform in color tone, making it cleaner overall, but it also looks slightly industrial and lacks a bit of breathability.

The control of information density could be refined further. Currently, the spacing between modules is too even, making it look more like a regular layout rather than a flow of content with emphasis, pauses, and guidance. The sense of undulation in the reading path is not obvious enough.

How much did it cost to complete all these tasks?

Finally, let's talk about the cost, which is what everyone is most concerned about.

This batch of cases covers SVG, Canvas, Three.js, games, security auditing, and long-running Next.js tasks. After running them all, I only used about 4% of my Kimi Code quota this week, which translates to approximately 6.99 yuan. For developers, the daily trial-and-error cost is already negligible.

Kimi K2.7 code excels particularly in front-end and interactive tasks. It can quickly build complete versions of SVG, Canvas, Three.js, and small games. For individual developers and tech bloggers, this...fastThe ability to validate ideas is extremely highpracticalvalue.

Even more exciting is that next Monday, Kimi will be releasing Kimi K2.7 Code at 6x speed..

The output speed is approximately 180 Tokens/s in typical programming scenarios and can reach 260 Tokens/s in short context scenarios.This will transform our development process from submitting requirements → long waits to extremely frequent, real-time interaction..

We and AI The collaboration paradigm is also clearer: we are responsible for defining the goals, checking the results, and making the final architectural decisions;AI Responsible for building the framework, writing the initial version, and making repeated revisions, completing a large amount of repetitive and mechanical work first.

AI They're already quite skilled at layout, color schemes, and page structure, and the generated pages are usually clean and well-organized. But the problem is that they're too organized; often, manual adjustments are still needed to transform the page from a template-like work into a truly well-designed and detailed piece.

open sourceThe competition among models has truly shifted from chasing ranking data to reducing costs and increasing efficiency within the industry.open sourceModel in programming and Agent In terms of tasks, it has begun to compete with top closed-source models, giving developers an unprecedentedly broad range of options in terms of architecture design, data security, and call costs.

Kimi K2.7 Code is just one example of this trend. With the 6x speedup version of Kimi K2.7 Code launching next Monday, will it be able to deliver both speed and stability? We'll have to wait and see.

Original link:Kimi K2.7 Code recreates Minecraft in less than 5 minutes – just how good is he?