Video summary

I Tested Every Kimi K3.1 Pro Mode for 24 Hours (vs Local K3 & DeepSeek) 🤯

Main summary

Key takeaways

Technology

Summary

The video tests Kimi K3.1 Pro in Standard, High, and Max thinking modes, comparing its cloud-based agent workflow with an earlier locally run Kimi K3 model. The reviewer evaluates generated interactive apps, games, simulations, and reasoning answers over roughly a day.

Product features and benchmark context

  • The reviewer describes Kimi K3.1 Pro as a 3-trillion-parameter model. Its chat interface includes agent-style tool use: it can work in a terminal-like environment, create files, run and inspect projects, take screenshots, and revise code.
  • The video cites a DeepSWE score of 75, compared with Claude’s 74. The reviewer cautions that DeepSeek V4.1 Flash’s higher result may rely on a special evaluation harness, making the comparison less direct.
  • Kimi’s interface also offers built-in examples and a 3D-generation workflow. Generated models can be downloaded in GLB format, and example prompts can be remixed.

Tests and observations

  • 3D solar system: Standard mode produced a polished, interactive model with clickable planets and a timeline. High mode used more tools and added refinements, such as a mobile-friendly interface. The reviewer observed that generated code can initially contain runtime errors, which the agent may fix through additional iterations.
  • Music-making app: Max mode created a browser-based, Logic Pro-inspired music tool. The reviewer was able to edit melodies and export audio as a WAV file, and considered this one of the strongest demonstrations.
  • Volcano simulation: Standard mode looked attractive but had excessive camera shake, and the lava initially did not flow correctly. After feedback, the lava behavior improved, though the camera remained rough. High mode had rendering and loading issues, along with visual artifacts. Max mode produced the strongest version, with gravity-driven lava, collapsing buildings, and a timeline. The result was still stylized rather than realistic.
  • Mars launch simulation: A remixed 3D scene showed a rocket launch, orbital travel, and a Mars arrival sequence. The reviewer noted odd behavior and visual inconsistencies during launch, but found the overall interactive sequence impressive.
  • Games: A simple bridge-running shooter generated in Standard mode included reinforcements, ammunition, and a final boss. An Age of Empires II-inspired RTS prototype supported basic resource gathering, farming, unit movement, and combat, but lacked some expected features, such as reliably creating new soldiers. The reviewer judged it a promising prototype that needed gameplay refinement.
  • Reasoning tests: The model struggled with the reviewer’s pronunciation and riddle prompts. Results varied by thinking mode and were ultimately considered inconclusive. In a coding-related question about Microsoft Store subscription product types, it found the relevant “durable” terminology after using web search.
  • Code quality: Generated projects can contain bugs, duplicated code, loading problems, or rendering artifacts. The video emphasizes that testing and follow-up prompts are often necessary, even when the first result looks convincing.

Local-model comparison and takeaway

  • The reviewer’s earlier local Kimi K3 setup ran at about six tokens per second and could produce useful results, but its output was less polished than the cloud demonstrations. A later local test was affected by a quantization issue.
  • The reviewer also describes splitting a roughly 580 GB model across three computers: two 128 GB machines and a 512 GB Mac Studio. They suggest local operation is possible, but much slower and dependent on suitable model quantization and hardware.
  • Overall, higher thinking modes generally used more tools and produced more refined results, with Max mode strongest in the volcano test. However, more reasoning did not guarantee flawless output, and Standard mode was already capable of producing useful interactive prototypes.

Reviews and guides covered

  1. Mode comparison: Standard vs. High vs. Max thinking, demonstrated through interactive app and simulation generation.
  2. Agent workflow overview: How Kimi uses tools, runs projects, inspects screenshots, and iterates on code.
  3. Local setup comparison: A high-memory, multi-computer setup for running a local Kimi model, with performance and quantization caveats.
  4. Generated-project review: Tests of a solar system, music app, volcano simulation, Mars mission, bridge shooter, and RTS prototype.
  5. Reasoning and coding checks: Riddles and a Microsoft Store product-type question, including the role of web search.

Main speaker/source: A single host and reviewer from the xCreate channel; the technical findings and opinions are based on the host’s demonstrations and tests.

Rate this summary

Your feedback will help improve summaries.

Improve this summary

Reprocess with a stronger model when the summary feels incomplete or inaccurate.

Pro

Translate summary in another language

Pro

Ask questions to this video

Chat for follow-up questions, clarifications, and source-backed answers.

Coming soon

Share this summary

Original video