Benchmarked K3 + Kimi Code on out-of-sample quant strategy generation — full report open-sourced

Hi team,

I run a Chinese-language channel on Bilibili/YouTube covering AI-assisted quantitative trading. Two of my recent videos used Kimi Code CLI as the primary driver.

Episode 1 (Jul 26) — K3 + Kimi Code writing an ETH perpetual futures strategy from scratch on Freqtrade. I imposed strict constraints: train/validation/test split, hard-capped stop-loss range, three loss functions cross-validated, and the model had to audit its own code for lookahead bias. K3 handled all of it and stayed within the rules — total spend ~¥8. That video became my highest view-to-follower conversion to date (29k views, 1,338 new subscribers).

Episode 2 (Aug 6) — a four-way benchmark: K3 (Kimi Code), Claude Opus 5 / Fable 5 (Claude Code), DeepSeek V4 Flash. Each model paired with its own official CLI, evaluated on physically isolated out-of-sample data (Jul 2025 – Jul 2026, benchmark index −45%), returns decomposed into alpha and beta. 116k views.

Everything open-sourced — configs, strategies, full reports:
https://github.com/frank-quant/ai-trading-videos

Two things:

Happy to share raw run logs if useful to the team — long-horizon agentic coding on a real toolchain surfaces things synthetic benchmarks don't.
I'll keep running these as new versions ship. If there's a channel for creator/developer relations, I'd like to connect.

Review the linked benchmark repository and the issue description first; no Kimi CLI file, test, or entry point is identified. Clarify whether the project is expected to publish documentation, review the report, or handle a creator-relations request, since no implementation scope or completion criterion is given.