Kimi K3 Open Source AI Beats GPT & Fable? 🤯 (Coding FF7, Red Dead & MLX)

Kimi K3 Open Source AI Beats GPT & Fable? 🤯 (Coding FF7, Red Dead & MLX)

by xCreate · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 xCreate 👥 26K 📅 July 18, 2026 ⏱ 22 min 👁 7K 📄 expert opinion 🧭 2026-09-16
Available in: English (current) Français

Keywords

Kimi K3open sourcecodingMLXgame development

Summary

The creator tests the newly released open weight Kimi K3 model by prompting it to create video games (Final Fantasy 7, Red Dead Redemption 2, GTA V, and a Rocky boxing game) and port an Inkling transformer model to Apple’s MLX framework. He highlights the model’s built-in ‘harness’ that autonomously iterates, taking screenshots and improving outputs, and its vision capabilities. Benchmarks show it competing with commercial models like GPT-5.6 Sol and Claude Fable 5. The video demonstrates impressive results, such as a fully playable FF7 clone with voice acting, but also notes high token consumption and occasional stubbornness in debugging. The MLX port succeeded after several iterations, fixing a simple indexing bug and using fused attention for speed. The creator concludes that while the model isn’t perfect, it significantly automates coding tasks.

132 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information lies in concrete demonstrations of the model’s capabilities, including complex game generation and MLX porting, which are rarely shown in detail. The argumentation is based on direct experience, and the creator honestly reports failures and limitations, such as the model getting stuck on a quantization hypothesis. However, the evaluation is subjective and lacks systematic benchmarking or comparison with other models under identical conditions. The claim that it beats GPT and Fable is based on third-party benchmarks mentioned but not elaborated. Overall, the argumentation is persuasive due to the visual evidence, but its scientific rigor is limited.

Scientific Rigor, Source Quality, Title Accuracy

The video cites no external scientific sources, relying instead on the model’s own benchmarks mentioned in the introduction. The description includes links to companion videos and tools like Inferencer, but these are not used as evidence. The title matches the content well, as the video primarily tests Kimi K3’s coding abilities and MLX porting. The lack of source citations undermines the scientific quality, but the hands-on approach provides practical insight.

186 words

Title / Content Match

The title accurately reflects the content: the video tests Kimi K3's capabilities in coding games and MLX porting, comparing it to other models.

Quality & Reliability

6/10

The video presents hands-on tests of Kimi K3 with detailed demonstrations, but relies on subjective experience and limited benchmark references. No external sources are cited beyond the model's own benchmarks. The porting effort is documented but not independently verified.

Chapters

Cited Sources

  • Kimi K2.7 — Companion video about a previous model iteration.
  • Kimi K2.6 — Companion video about an earlier model iteration.
  • Inferencer App — Tool used for MLX testing.

Concurring Sources

  • Kimi K2.7 — Similar testing of a previous Kimi model, showing evolution.
  • Kimi K2.6 — Earlier model test, providing baseline for comparison.

External References

Contribution & Novelties

The video’s original contribution is a hands-on exploration of the newly released Kimi K3 model, demonstrating its autonomous coding and debugging capabilities through complex game creation and an MLX port. It reveals both strengths (self-correcting, vision, audio generation) and weaknesses (token consumption, occasional stubbornness). This provides practical insights for developers considering open weight models.

Pour aller plus loin :

  • MLX (Apple’s machine learning framework) — Relevant for understanding the porting target.
  • Quantization (machine learning) — Explained the 3-bit quantization issue encountered during debugging.
  • Autonomous agent — Related to the model’s built-in ‘harness’ that iteratively improves its output.

97 words

Radar Profile

The radar profile shows high scores for information quantity and technical level, with moderate quality and reliability. This indicates a content-rich but not fully verified presentation, typical of an expert hands-on review.

Reliability 6/10