Kimi K3.1 Pro: A Deep Dive into its Frontier-Class Open-Weights Capabilities

Kimi K3.1 has been launched as a frontier-class open-weights model featuring 3 trillion parameters. The reviewer acquired a subscription, downloaded the Kimi desktop app, and began building various demonstrations to showcase its capabilities.

The review will delve into benchmarks, multiple demos, and model specifications, asserting that the model demonstrates high performance.

The Kimi K3.1 model underwent a 3D Rubik's Cube simulation test, a task where many open-weights models typically falter due to issues with 3D physics, color consistency, and rotation logic. In contrast, Kimi K3.1 delivered a hyperrealistic and accurate 3D rendering of the Rubik's Cube, complete with real-time lighting reflections, demonstrating its proficiency in complex 3D simulations.

The model includes built-in controls for scrambling, solving, and adjusting rotation speed, allowing for interactive manipulation of the cube. Importantly, it employs an actual algorithm to solve the cube, rather than merely reversing moves, indicating a deeper understanding of the task.

The reviewer noted the presence of "eyebrow text," a characteristic often seen in Claude family models, and mentioned accusations of distillation attacks against Kimi, though this topic was deferred for future discussion. The simulation also featured keyboard controls and a timer for quick solving.

Kimi K3.1 demonstrates strong 3D modeling and animation proficiency, joining other leading models such as Opus 5.5 and GPT 6.1 Soul in this area. This convergence suggests a growing industry focus on advanced 3D capabilities within AI models.

Google's Argon model also handles 3D modeling reasonably well, though the reviewer judged it weaker than Kimi, Opus, and GPT 6.1 Soul. Meta's Muse Wonderful is described as a phenomenal model overall, but it is not good at 3D modeling. Alexander Wang, head of AI at Meta, acknowledged the gap, saying, "Oops, we kind of forgot to make it good at 3D modeling. We should have probably done that."

The reviewer emphasized the importance of 3D modeling as a testing method, often using it to build games or 3D models to evaluate a model's performance.

Kimi K3.1 was tasked with creating a 3D WebGL rendering of the 'new iPhone duo,' with instructions to aim for accuracy and realism in a browser-based format. The model produced a visually impressive rendering that included realistic lighting and reflections off the metallic edges and camera bezels, showcasing its high-quality visual output.

However, the rendering suffered from significant structural inaccuracies. The Apple logo appeared incorrect, and the unified screen of the iPhone duo was depicted with a large split down the middle, deviations from the actual device's appearance. This indicated a gap between the model's rendering capabilities and its factual accuracy.

The generation process for this 3D rendering took 28 minutes and 45 seconds. Despite explicitly searching the web for information during this time, the model's output remained inaccurate. The reviewer initiated a follow-up prompt, instructing Kimi K3.1 to "Research what the iPhone duo actually looks like, the actual specs and measurements, and recreate it again" to encourage a more accurate iteration.

The reviewer noted that Kimi K3.1 Pro was being used on "high reasoning effort" for this task, rather than "max reasoning effort," to maintain consistency in testing parameters.

The model generated a 28-second, 1080p motion graphic designed to explain how to tie a figure-8 knot for belaying. The graphic covered steps such as forming the figure-8, threading tie-in points, retracing the knot, dressing and tightening, and checking tail length.

Despite its attempt, the motion graphic lacked visual clarity, particularly regarding how different segments of the rope overlap, which is crucial for understanding knot tying. The reviewer concluded that the output was insufficient for actually learning how to tie the knot. The model also included a safety warning, advising that it is "not a substitute for instruction," similar to warnings observed in ChatGPT and Claude models.

The reviewer compared Kimi K3.1's motion graphic for tying a belaying knot with one generated by ChatGPT-6.1 Soul. While Kimi's output was visually unclear regarding rope overlaps, the ChatGPT version provided a much clearer visual representation of the rope's path, making it easier to follow the knot-tying process.

The reviewer stated that if they had a rope and were following the ChatGPT motion graphic, they would "absolutely be able to tie this knot," highlighting ChatGPT's superior instructional clarity. Both models included safety warnings, advising users to learn and practice with a qualified instructor, a similarity the reviewer noted as interesting, suggesting Kimi behaves similarly to ChatGPT and Claude models.

Following this comparison, the reviewer prompted Kimi again, asking it to make the motion graphic "prettier" with a "premium climbing brand film" aesthetic, more tangible rope, animated knot, continuous camera movement, simplified typography, and refined harness illustration. This iterative attempt aimed to improve the aesthetic and functional aspects of the graphic.

Following a prompt to enhance the visual appeal of the belaying knot motion graphic with a 'premium climbing brand' aesthetic, Kimi K3.1 produced a second version. This iteration incorporated drop shadows and improved camera movement, giving the rope a more 3D appearance.

Despite these aesthetic refinements, the model still failed to clearly illustrate when one part of the rope overlapped or went under another, a critical detail for instructional accuracy. The reviewer noted that while there was more movement and a clearer distinction between the two rope pieces, it was still not good enough to learn how to tie the knot.

The model's performance in this task was deemed decent but "certainly nowhere near what the frontier is capable of," indicating that despite improvements, it remains below top-tier standards for this type of detailed instructional content.

For a new test, the reviewer tasked the model with building a 'Fall Guys' clone for the browser, instructing it to reference the original game via web search for videos and images. The model generated a 3D obstacle race called 'Bean Blitz' in 48 minutes, featuring 12 animated CPU rivals, various mechanics such as rotating sweepers, sliding gates, bouncing pads, and swinging balls, as well as customizable character traits, sound effects, and adjustable game parameters.

The reviewer emphasizes the effectiveness of keeping prompts "nice and short" for open-weights models. This approach, he explains, allows him to better understand the model's inherent capabilities and its "taste" by enabling it to make more assumptions about user intent. Conversely, detailed prompts are recommended only when specific aesthetic or technical constraints, such as particular colors, animations, designs, or level elements, are required, or when specifying the technology to be used.

The generated 'Fall Guys' clone, 'Bean Blitz,' was critiqued for its visual and technical performance. The reviewer noted that the colors appeared "too pastelly," and there were issues with "inaccurate collisions" and "a lot of flickering."

Despite these flaws, the reviewer acknowledged that the model did "a decent job, especially for an open weights model," expressing appreciation for open-weights, open-source models in general. The initial version of the game also lacked music.

Kimi K3.1 was tasked with redesigning the forwardfuture.com website. The model successfully adhered to the provided brand guidelines, utilizing the correct colors and fonts. The reviewer noted that the redesigned version "looks pretty good" in terms of its general aesthetic.

However, the design exhibited typical "AI traits" and included "incorrect spacing," making it feel like "AI design" rather than a human-crafted one. The reviewer commented that it "feels like what ChatGPT would use," suggesting a generic or predictable quality. Overall, the result was functional but lacked distinctiveness.

The reviewer generated a 5-minute YTP (YouTube Poop) parody video of Theo, using Kimi K3.1. Curious about the underlying process, the reviewer asked Kimi how it created the video, revealing an unexpected technical pipeline. Kimi explained that it utilized an "AI video generation" API, specifically "Kimmy's video generation API configured as V2," to create five original animated parody clips.

The process involved Python for orchestrating the edit and generating the procedural electronic soundtrack with NumPy. FFmpeg was used for various video manipulations, including reversals, mirroring, color distortion, captions, clip sequencing, and final encoding. FFprobe was employed to verify the 5-minute runtime and the actual video format.

The reviewer expressed impressiveness with the technical output, despite the video's intentionally "garbage" content. He noted, however, that he frequently has to "nudge Kimmy to do things" more so than other models that might make assumptions about user intent, indicating a need for explicit instruction with Kimi.

When tasked with creating a graph displaying the benchmarks of Kimi K3.1 Pro against GPT 6.1 Soul, Opus 5.5, and DeepSeek 4.1 flash 10 18, Kimi initially provided only broad categories such as coding, reasoning, mathematics, and agentic tasks, without specific benchmark names.

Upon being prompted to list specific benchmarks, Kimi then hallucinated inaccurate placeholder numbers for various benchmarks like SweBench, verified terminal bench, GPQA Diamond, and AMIE 2026. The reviewer had to explicitly ask if these numbers were accurate, to which Kimi responded, "No, they are invented placeholders."

To obtain a usable output, the reviewer had to manually provide the correct published benchmarks to Kimi. Even then, Kimi assembled the data but did not recreate the image, requiring another explicit "nudge" to generate the graph. The final image largely recreated the one published by Kimiko with colorization. Based on these tests, the reviewer concluded that Kimi K3.1 is a very good model but not yet on par with top-tier competitors like Opus 5.5 and GPT 6.1 Soul, suggesting a potential underlying connection to those models due to similar behaviors.

The reviewer highlighted the need for frequent 'nudging' for Kimi K3.1 to perform tasks correctly, as it often fails to make appropriate assumptions about user intent, unlike other models.

Kimi K3.1 Pro is confirmed as a 3 trillion parameter class open-weights model. The reviewer expressed strong appreciation for Kimiko for releasing "phenomenal massive models for free."

The availability of such models allows users to download and run them on their own data centers, directly use Kimi, or access them via Neo clouds, fostering innovation within the open-source community. The reviewer emphasizes the significance of these open-source, open-weights models becoming increasingly competitive with frontier models.

Kimi K3.1 Pro boasts a 1 million token context window and a total of 2.8 trillion parameters, with 104 billion active parameters. This configuration indicates it is a Mixture of Experts (MoE) model, where only a portion of parameters are activated per prompt, optimizing efficiency.

The model supports multimodal input, definitely accepting images and likely audio. It specializes in full-project coding for game engines like Unreal, Unity, and Godot, as well as parametric 3D modeling in Blender and CAD. Kimi K3.1 is also noted for its reliability in finance tasks and a reported 65% reduction in hallucination rates. It aims to match proprietary models in coding, agentic, and knowledge benchmarks, though the reviewer suggests caution regarding the term "benchmarks" given previous observations of data inaccuracies.

In the Terminal Bench, which tests a model's ability to write and execute terminal commands, Kimi K3.1 Pro scored 65.7, placing it just under Opus 5.5's 66.4. In the Sweep Marathon benchmark, Kimi outperformed Opus 5.5, scoring 51.7 compared to Opus's 46.3.

On OS World, a benchmark that measures a model's ability to control a computer, Kimi K3.1 scored a strong 79.6, in line with Opus and Fable. The source notes that GPT-6 Astro placed fourth on the same benchmark, and the reviewer says he personally considers it the best at computer and browser use. On Terminal Bench Science, however, Kimi scored about half of GPT-6 Astro's result. On GDP Val, an OpenAI evaluation of real-world knowledge work, Kimi K3.1 ranked second with an Elo score of 1846, just behind Opus 5.5 at 1866.

The reviewer also noted Kimi's strong showing in Pareto curves, which aim for the highest score, highest quality, and lowest price, positioning Kimi in a favorable spot. However, the reviewer mentioned that despite GPT-6 Astro placing fourth on the OS World benchmark, he personally finds it to be the best for computer and browser use.

Kimi K3.1 Pro's regular speed version is priced at $2 per million tokens for input and $10 per million tokens for output. This pricing is considered "very competitive" when compared against similarly capable models, offering a high-performance option at an attractive cost. Cached input tokens are even more affordable at $0.20 per million.

For a high-speed version, the input and output prices double to $4 and $20 per million tokens, respectively, while the cached input price remains at $0.20. This high-speed option is noted to be two to three times faster, providing a significant performance boost for a relatively modest increase in cost, making it a compelling choice for users prioritizing speed.

Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, placing it ahead of models like Fable 5.1, GPT-6 Astra Max, and Argon, and directly behind Opus 5.5, Sonnet 5.5, and Muse Wonderful. This indicates that, according to these benchmarks, Kimi K3.1 Pro is considered an "absolute frontier model" while being open source and open weights.

Regarding the cost per intelligence task, a metric that accounts for both token usage and token cost, Kimi K3.1 Pro stands out with a cost of $2.11. This is notably lower than other top models, such as Yimou's Wonder Max at $4.54 and GPT-6 Astra, demonstrating its significant cost-efficiency in completing tasks, making it a highly attractive option in the market.