Kimi K3.1
A next-generation open-source model from Moonshot AI currently in pre-release preparation. It is designed to address known limitations of K3, with the goal of matching or surpassing top-tier models in the Claude and GPT series while reducing token pricing.
Core Facts
On Hacker News this partnership topic received 90 points and 8 comments.
In a Minecraft sandbox game generation comparison, Kimi K3 took 2.5 hours and cost $35, while DeepSeek V4.1 Flash 1017 took 1.5 hours and cost 80 cents, nearly 44 times cheaper.
Kimi K3.1 scored 55 on the Artificial Analysis composite benchmark.
Timeline (last 90 days)
Kimi K3.1 reportedly appeared in an API request and was shut down 7 minutes later; leaked information shows it uses the same tokenizer and chat format as Kimi K3 but introduces 4 reasoning levels (low, medium, high, and highest, with highest as default) and supports image input.
The Fall Guys clone game "Bean Blitz" generated by Kimi K3.1 took 48 minutes and includes 12 animated characters and complete functionality, but has issues such as a pastel-leaning color scheme, inaccurate collision detection, and screen flickering.
The 54-second rope animation produced by GPT 6.1 Sol at medium reasoning effort clearly shows the rope passing over and under, whereas neither version of Kimi K3.1's animation makes the rope's occlusion and interleaving relationships recognizable.
In the test of modeling a climbing protection knot in Blender, Kimi K3.1 failed: the knot shape was close to a figure-eight knot but was not actually tied, and it was not connected to the harness.
In the "render an iPhone Duo with WebGL" task, Kimi took 28 minutes 45 seconds to complete, with decent rendering quality but a completely wrong product form (a deformed Apple logo, missing the crease design).
In Berman's 3D Rubik's cube test, Kimi K3.1 fully rendered a 3D Rubik's cube with lighting reflections and implemented keyboard control, a timer, undo and reset, rotation speed adjustment, and algorithm-based automatic solving, earning the assessment of "an absolute pass, and one of the best."
Tech blogger Matthew Berman, after hands-on testing, commented that Kimi K3.1 is a great model but still has a distance to go before Opus 5.5 and GPT 6.1 Sol.
The official blog claims that Kimi K3.1 has improved capabilities on project-level coding, parametric 3D modeling, and financial modeling tasks, with the hallucination rate down 65%.
Kimi K3.1's model weights are downloadable, can be deployed in one's own data center, and can also be called directly through the Kimi platform.
Kimi K3.1 lies on the Pareto frontier of the Artificial Analysis leaderboard.
23 more timeline events
All Facts (20)
On Hacker News this partnership topic received 90 points and 8 comments.
65%Partially VerifiedIn a Minecraft sandbox game generation comparison, Kimi K3 took 2.5 hours and cost $35, while DeepSeek V4.1 Flash 1017 took 1.5 hours and cost 80 cents, nearly 44 times cheaper.
70%Partially VerifiedKimi K3.1 scored 55 on the Artificial Analysis composite benchmark.
65%UnverifiedKimi K3.1 reportedly appeared in an API request and was shut down 7 minutes later; leaked information shows it uses the same tokenizer and chat format as Kimi K3 but introduces 4 reasoning levels (low, medium, high, and highest, with highest as default) and supports image input.
50%UnverifiedTech blogger Matthew Berman, after hands-on testing, commented that Kimi K3.1 is a great model but still has a distance to go before Opus 5.5 and GPT 6.1 Sol.
50%UnverifiedKimi K3.1's model weights are downloadable, can be deployed in one's own data center, and can also be called directly through the Kimi platform.
50%UnverifiedThe official blog claims that Kimi K3.1 has improved capabilities on project-level coding, parametric 3D modeling, and financial modeling tasks, with the hallucination rate down 65%.
50%UnverifiedKimi K3.1 lies on the Pareto frontier of the Artificial Analysis leaderboard.
50%UnverifiedIn the benchmark comparison charts shown in the livestream, K3.1 sits close behind Opus 5.5 and GPT-6 Astra Fable 5.1 on metrics such as Terminal Bench, Humanity's Last Exam and GPQA, and surpasses them on some items.
50%UnverifiedKimi K3.1 uses a MoE architecture, with about 2.8 trillion total parameters and 104 billion activated parameters.
50%UnverifiedKimi K3.1 supports a one-million-token context, is natively multimodal, and can directly output code-driven game projects, covering the three major engines Unreal, Unity and Godot.
50%UnverifiedAccording to the official explanation shown in the livestream, K3.1's model hallucination rate on financial tasks dropped by about 65%.
50%UnverifiedIn a climbing-knot animation comparison, the output of GPT-6.1 (6.1 Sole Medium) was about 10% to 15% better than K3.1 in spatial relationships.
50%UnverifiedAccording to logs read out in the livestream, Kimi's video generation call path goes through a Python FFmpeg processing pipeline.
50%UnverifiedWhen releasing K3.1, Moonshot did almost no promotion of its video generation capability, but the actual interface is already available.
50%UnverifiedMatt pointed out that the horizontal-axis labels on one of the benchmark comparison charts were generic names rather than specific benchmark names, and that Kimi K3.1 ranked first on every item, questioning whether the chart was self-made.
50%UnverifiedKimi K3.1 Pro retains a 1M token context window and image input capability.
50%UnverifiedKimi K3.1 Pro is the first model in the Kimi K3.1 series.
50%UnverifiedThe 54-second rope animation produced by GPT 6.1 Sol at medium reasoning effort clearly shows the rope passing over and under, whereas neither version of Kimi K3.1's animation makes the rope's occlusion and interleaving relationships recognizable.
50%UnverifiedThe Fall Guys clone game "Bean Blitz" generated by Kimi K3.1 took 48 minutes and includes 12 animated characters and complete functionality, but has issues such as a pastel-leaning color scheme, inaccurate collision detection, and screen flickering.
50%