Published: 2026-06-14

Kimi K2.7 vs GLM-5.2 in Hermes Agent: Real Coding Showdown

Chapters / key moments (click to jump — plays here on the page)

Fahd Mirza runs two of China's strongest open coding models — Moonshot's Kimi K2.7 Code and Zhipu's GLM-5.2 — head-to-head inside Hermes Agent on the same Ubuntu box. The job isn't a toy: a real World Cup standings app with hundreds of files and a deliberately planted bug. Both models have to find the bug, fix the root cause, and build a brand-new round-of-32 bracket feature in one shot, then face a from-scratch HTML generation test.

Source video

"Kimi K2.7 vs GLM-5.2: Real Coding Showdown in Hermes Agent" by Fahd Mirza — Watch on YouTube →

The two models

Kimi K2.7 Code (Moonshot AI) — a 1-trillion-parameter mixture-of-experts model with ~32B active, a 256K context window, built for agentic coding with thinking mode always on.
GLM-5.2 (Zhipu) — ~744B parameters, a full 1-million-token context, shipped under MIT, with open weights described as arriving the following week.

Key Takeaways

The bug: the app ranked third-place teams by points alone, ignoring goal difference — so Ghana (better goal difference) was wrongly knocked out below Ecuador despite equal points. Correct FIFA tie-breaking should advance Ghana.
GLM-5.2 fixed the root cause correctly (~97 tool calls), re-ordered the three-point teams by goal difference, and built a full 16-match round-of-32 bracket with group winners seeded and no team facing a rival from its own group.
Kimi K2.7 produced the same correct fix and bracket in less time (~5 minutes) with more tool calls, and added an extra "future progression" touch on its own initiative.
Second test (a single self-contained HTML file simulating a Siberian cold wind from Lake Baikal to Pakistan with an animated particle system and live map): GLM-5.2's version looked clearly better, with some real animation; Kimi's was flatter and lacked visible simulation.
Verdict: neck-and-neck. Kimi had a slight edge on the agentic bug-fix-and-build task (speed + initiative); GLM-5.2 won the creative one-shot generation. Both run inside Hermes Agent today by setting the provider and model. Test on your own use case before deciding.

Commands & Code Mentioned

# Inside Hermes Agent: set the provider and model, then launch
# (GLM-5.2 in one terminal tab, Kimi K2.7 in another)
hermes set provider/model   # set provider + model for each run
hermes                      # launch the Hermes Agent session

# Run the buggy app under a fresh virtualenv to verify the fix
python -m venv .venv && source .venv/bin/activate
# seed data + run; app serves on http://localhost:5000

Exact provider/model selection syntax depends on your Hermes config — see the best free models for Hermes and Hermes setup guides.