Kimi K3: An Open Source #1 — Coding Tests vs Fable 5, Opus 4.8 & GPT-5.6 Sol

Overview

Kimi K3 is Moonshot AI's new 2.8-trillion-parameter open model — the first open model ever in the 3T class. With a 1-million-token context window, native multimodal input (text, image, and video), and performance that rivals the best closed models, K3 represents what Alejandro calls "a DeepSeek moment" for open-source AI.

Demand was so high at launch that Kimi had to pause new subscriptions. In this video, Alejandro puts K3 through its paces with real coding benchmarks, his own GitHub issue tests, and head-to-head comparisons against Claude Fable 5, Claude Opus 4.8, and GPT-5.6 Sol.

Specs & Architecture

K3 is the first open model in the 3-trillion parameter class. It features native multimodal capabilities across text, image, and video inputs, and its 1M-token context window makes it suitable for processing entire codebases. Open weights are releasing on July 27, enabling the community to fine-tune and deploy.

Design Arena: First Open Model at #1

K3 tops the Design Arena frontend coding leaderboard — making it the first open model ever to hold the #1 position. This leaderboard measures real-world frontend and webdev coding ability through human preference voting, not synthetic benchmarks.

Artificial Analysis Intelligence Index

K3 sits at the top of the Artificial Analysis Intelligence Index among open models. This index aggregates performance across multiple benchmarks into a single composite score, providing a holistic view of model capability. K3 doesn't just lead open models — it competes directly with closed frontier models.

Terminal-Bench 2.1 & Token Usage

On Terminal-Bench 2.1 — which tests real terminal and coding tasks — K3 goes toe-to-toe with Claude Fable 5 and GPT-5.6 Sol. The benchmark includes complex multi-step terminal workflows that mirror real developer tasks.

Token usage analysis shows K3 is competitive in efficiency compared to frontier closed models, despite its massive parameter count. The model doesn't waste tokens on unnecessary verbosity — a common issue with large MoE models.

Real-World GitHub Issue Tests

Alejandro runs his own benchmark on a real GitHub issue, comparing quality, cost, time per task, and token usage across four frontier models. This goes beyond synthetic benchmarks to test real-world problem-solving.

Benchmark Comparison

K3 delivers competitive quality at significantly lower cost — the key advantage of open models with efficient inference.

Run Your Own Evals with DuoBench

DuoBench is Alejandro's open-source tool that lets you run the same evaluations on your own repositories. It enables reproducible benchmarking for your specific use cases — no more relying solely on generic leaderboards.

Using Kimi K3 in Pi, Tau & OpenCode

Alejandro shows how to start using K3 today in three coding tools:

All three tools provide terminal-based coding interfaces where K3 can be used as the backing model. Setup involves pointing the tool's API config to the Kimi platform endpoint and selecting K3 as the model.

Key Takeaways

Chapters

Resources & Links