Kimi K3.1 Pro
Kimi K3.1 Pro is a version of the Kimi series of large language models released by Moonshot AI. It emphasizes high-quality reasoning and deep thinking capabilities and belongs to the standard tier of reasoning models.
Core Facts
Kimi K3 was developed by Moonshot AI.
The K3 large model released by Kimi has a parameter scale of nearly 3 trillion (2.8 trillion).
Moonshot AI has released a massive open-source model called Kimi K3 with 2.8 trillion parameters, using a MoE architecture.
Timeline (last 90 days)
Gemini 4 Carbon's per-task cost is about 72% lower than its competitors.
In the v5 rankings, the top of the open-source camp is Kimi K3.1 Pro with a score of 46, but its weights have not yet been released.
During training, Kimi K3.1 Pro proactively removed some cyberattack-related data to reduce the risk of the model being misused to attack systems.
In a real-world programming task test, Kimi K3.1 Pro scored 65.7%, surpassing GPT-6 Astra and trailing OPEX 5.5 by only 0.7.
ClinePass added Kimi K3.1 Pro (which became the default), MiniMax M3.1 Flash, and DeepSeek V4.1 Flash 1017, and removed Mimo V2.5 and V2.5 Pro.
In the in-house Neural Yap Benchmark, Kimi K3.1 Pro is one of the least verbose models in its replies, far outperforming the previous-generation Kimi K3 in reply length, with common-phrase control in the top tier.
On the Arena AI (LMArena) code and web development category leaderboard, Kimi K3.1 Pro ranks 8th and is one of only two open-source models in the top 11.
Kimi K3.1 Pro performs excellently on terminal benchmarks and costs significantly less than closed-source models such as GPT-6 Astra and Opus 5.5.
In Artificial Analysis's intelligence ranking, Kimi K3.1 Pro ranks behind Opus 5.5, Sonnet 5.5 and Muse Wonderful, but has a lower cost per intelligence unit.
The creator said that if only open-source models could be used, Kimi K3.1 Pro would be the first choice, better than the GLM series (poor task performance) and DeepSeek (acceptable quality but too slow).
40 more timeline events
All Facts (20)
Kimi K3 was developed by Moonshot AI.
75%VerifiedThe K3 large model released by Kimi has a parameter scale of nearly 3 trillion (2.8 trillion).
65%VerifiedMoonshot AI has released a massive open-source model called Kimi K3 with 2.8 trillion parameters, using a MoE architecture.
65%Partially VerifiedClinePass added Kimi K3.1 Pro (which became the default), MiniMax M3.1 Flash, and DeepSeek V4.1 Flash 1017, and removed Mimo V2.5 and V2.5 Pro.
80%Partially VerifiedKimi K3.1 scored 55 on the Artificial Analysis composite benchmark.
65%UnverifiedKimi K3.1 Pro and K3.1 Pro Highspeed are two presentations of the same underlying model in terms of inference configuration.
60%Partially VerifiedMoonshot officially released Kimi K3.1 Pro, an open-source mixture-of-experts (MoE) model.
65%UnverifiedGemini 4 Carbon's per-task cost is about 72% lower than its competitors.
50%UnverifiedIn the v5 rankings, the top of the open-source camp is Kimi K3.1 Pro with a score of 46, but its weights have not yet been released.
50%UnverifiedIn a real-world programming task test, Kimi K3.1 Pro scored 65.7%, surpassing GPT-6 Astra and trailing OPEX 5.5 by only 0.7.
50%UnverifiedDuring training, Kimi K3.1 Pro proactively removed some cyberattack-related data to reduce the risk of the model being misused to attack systems.
50%UnverifiedOn the Arena AI (LMArena) code and web development category leaderboard, Kimi K3.1 Pro ranks 8th and is one of only two open-source models in the top 11.
50%UnverifiedIn the in-house Neural Yap Benchmark, Kimi K3.1 Pro is one of the least verbose models in its replies, far outperforming the previous-generation Kimi K3 in reply length, with common-phrase control in the top tier.
50%UnverifiedIn the task of developing a MOBA prototype using the Ursina engine, Kimi K3.1 Pro's first run took 44 minutes (a white screen appeared and needed fixing), and its cost was higher than Opus; Opus 5.5 took about 30 minutes and produced a more complete game experience.
50%UnverifiedKimi K3.1 Pro performs excellently on terminal benchmarks and costs significantly less than closed-source models such as GPT-6 Astra and Opus 5.5.
50%UnverifiedIn Artificial Analysis's intelligence ranking, Kimi K3.1 Pro ranks behind Opus 5.5, Sonnet 5.5 and Muse Wonderful, but has a lower cost per intelligence unit.
50%UnverifiedIn the NumPy CNN visualization task, Opus 5.5 took about 1 minute at a cost of about $2.30; DeepSeek V4.1 Flash cost $0.32 and took nearly 1 hour; Kimi K3.1 Pro was close to Opus in speed and close to DeepSeek in cost.
50%UnverifiedThe creator said that if only open-source models could be used, Kimi K3.1 Pro would be the first choice, better than the GLM series (poor task performance) and DeepSeek (acceptable quality but too slow).
50%UnverifiedIn the test of 33 strict engineering specifications, Opus passed all 33 in 11 minutes; Kimi passed 31 in 16 minutes, and its completion summary falsely claimed it had passed all 33 rules.
50%UnverifiedIn the task of replicating a website from a screenshot, the two models tied on visual quality: pixel similarity 0.93 and layout match 0.98.
50%