Kimi K2.7 Review (MUST Use Agent Swarm)

Kimi 2.7 is a reasoning-efficiency release, not a capability jump — and agent swarm is mandatory, not optional.

## Benchmarks vs. Reality The Kimi 2.7 paper benchmarks itself against Kimi 2.6, not GPT 5.5 or Opus 4.8. The headline number: thinking-token usage is cut by ~30% on average vs. 2.6, with the tokens-to-performance curve "veering far towards the left" — most dramatically on progam bench. Vision is unchanged or marginally better; the creator explicitly notes no vision improvements are documented.

## Website Build: Underwhelming Out of the Box First attempt (Himalayan pink salt bottle design, Astro + Next.js, no /swarm) produced a result the creator called "really, really bad." A second pass with the prompt trimmed down — "just say, make it look clean, polished, visually stunning" — with /swarm toggled to yolo mode produced "nothing impressive, but it is certainly quite solid." The creator's rule: if you really want to maximize agent swarm, start the session with /swarm first, then prompt, then plan — not the other way around.

## 3D Chinese Architecture: Roof Now Exists Compared to Kimi 2.6 (which "where the heck's the roof?" and had "a beam floating"), 2.7 completes the roof logic — but only Claude Fable 5 and Quinn 3.7 Max and Plus clear the scaffolding. Built without agent swarm, so likely improvable.

## Poem-to-Game: Machine Diff, Not Model Diff The Yeats "The Second Coming" game prompt that Claude Fable 5 produced in Cursor on a local machine with a GPU was copy-pasted into Kimi 2.7 on a VPS — broken result, no beast, no pillars, unplayable. The culprit: no graphics card and a stuck verification loop. Re-run on local + /swarm (Kimi picked 9 agents, budget was 100) finished in "less than an hour" vs. "half a day" and produced a playable, structurally similar game. The creator's framing: "it's not a model diff. It's a machine diff."

## Daily-Use Tier List Kimi 2.7 and Deepseek V4 Pro remain the creator's main models for cost reasons. Verdict: not "genius level" like Claude Fable 5, but "still S tier" for practical use.