Kimi K3.1 Pro Hands-On: China's Strongest Open-Source LLM Tested for 2 Days – Surprises & Complaints

Introduction

On Programmer's Day (1024), Kimi released K3.1 Pro as an open-source model claimed to be the closest to top-tier closed-source models. The author, an AI evaluator, subscribed to the Pro tier (¥199) and spent two days testing its capabilities across multiple domains.

1. 3D Modeling & Visualization

Cow Dissection for Education

The author tasked Kimi with creating an interactive 3D cow stomach model for a child's homework. Running in Kimi Code, the task took about 10 hours and consumed 262 million tokens (254 million cache hits). Most time was spent on asset acquisition, conversion, 3D design, and complex interactions. The resulting model shows the four stomach chambers (rumen, reticulum, omasum, abomasum) but uses schematic rather than high-fidelity geometry.

Yellow Crane Tower & 3D Printing

Using the web version's "3D Make" mode, the author generated three artifacts for the Yellow Crane Tower: a colored GLB for 3D preview, a white 3MF for single-color slicing, and a high-detail assembly preview with exploded view and 21-step assembly animation. The process required iterative refinement; feeding a VR panorama helped achieve satisfactory results.

Singapore Changi Airport in Blender

Kimi K3.1 Pro also handled a complex Blender project replicating Changi Airport, completing asset creation, scene assembly, animation, and rendering end-to-end.

2. Document-to-Video Generation

Gradient Descent Animation

Prompt:

创建一个展示梯度下降过程的 3D 动画。使用一系列路径点，模拟一个小球从波峰逐步向波谷移动、不断逼近最优解的过程。同时，介绍梯度下降属于哪一类机器学习算法，并列举其主要应用场景

The output shows a ball oscillating and converging on a loss surface with concept cards and application notes naturally integrated. Compared to Claude Opus 5.5 (considered the current quality ceiling), Opus 5.5 edges out in camera rhythm and lighting detail, but Kimi's result is highly usable for teaching.

Technical Article to Video Pipeline

The author built a multi-step pipeline: fetch Kimi K3.1 Pro technical blog via a custom skill (z-web-pack), extract script and cloned voice (z-article-to-video), then generate Remotion code aligning narration, subtitles, and animations. A flowchart illustrates the complexity. The final video suffered from slightly slow pacing, requiring further debugging. The author notes that any single-step error (subtitle timing, chart cropping, truncation) appears in the final cut.

3. Game Development

Given a vague prompt ("make a cycling game demo with familiar stack"), Kimi chose vanilla JavaScript + Three.js + Vite and delivered a smooth playable demo. The author considers this level of task relatively simple for a large-parameter model. Additional games (Merge Watermelon, FPS, Pelican Speed Cycling) were also produced with good completeness.

4. PPT & Interactive Webpages

Kimi's long-standing strength in PPT design persists. The author generated an "Embodied Intelligence Industry Research Report" PPT and an online interactive version requiring almost no edits. For a more complex request ("generate LLM industry knowledge graph covering upstream/midstream/downstream"), Kimi produced a graph with 104 nodes, 204 relationships, and 9 segments covering chips, data, compute, model R&D, training/alignment, platform toolchains, applications, industry, and terminal ecology, with search, three layouts, node details, and shortest-path tracing.

Benchmark Comparison & Capability Analysis

Official benchmarks show Kimi K3.1 Pro leading among Chinese models and open-source models on core programming benchmarks, even surpassing GPT-6 Astra. The author cross-referenced Kimi's technical blog with published scores for Claude Opus 5.5, GPT-6 Astra, and DeepSeek-V4.1-Flash-1017, taking the intersection of test items to reduce cherry-picking. The aligned results indicate K3.1 Pro still trails the strongest models (Claude Fable 5, GPT-6 Sol) on high-complexity pure logic reasoning and comprehensive ceiling tasks.

Strengths: 3D modeling & visualization, agent workflows, ultra-long-context real-world software engineering, cross-tool debugging (evident in the video pipeline).

Weaknesses: 1) Pure "brain-only" math/logic puzzles without tools; 2) Vertical domain professional research-grade computation; 3) Offensive cybersecurity development and vulnerability penetration (active safety downgraded).

Pricing Analysis

K3.1 Pro improves over K3 while being slightly cheaper. A Highspeed variant offers 3× speed at 1.5–2× cost. The author warns against Highspeed for non-critical use: it consumes tokens extremely fast, exhausting the 5-hour allowance quickly, and the extra cost is not justified for marginal speed gains. Overall, the author criticizes Kimi's package pricing as unreasonable — equivalent to a Claude Pro or ChatGPT Plus subscription — and suggests lower pricing would reduce negative sentiment.

Conclusion

Kimi K3.1 demonstrates that Chinese open-source models are rapidly closing the gap with global top-tier closed-source models through real-world combat testing. The broader Chinese LLM ecosystem is climbing the peak via diverse approaches, and the era rewards those tackling the hardest problems.