Kimi K3 (fast router) in OpenCode 1.18.35
skills on · OpenCode 1.18.35 + skills on (54: 33 video + 21 web/Three.js)Kit: 33 video + 21 web/Three.js · 1 of 54 installed skills used in the session
The model's own page, run in a sandboxed frame only when you ask (a page that needs its own origin opens in a new tab on a separate demo site). Phones get a full-screen player capped at 1.5x pixel density (2x on larger screens) so it stays smooth; the original page runs uncapped. It starts in its auto-tour (T switches to play mode on a keyboard).
Wall time 45m 01sCost $3.00 reported
Review
No verdict yet.
The full review
Stopped by us. Stopped by us (box row cap reached (2700s); stopped by the box lane so the run ships as partial); not a model failure.
Scores
Not scored: only runs that finished on their own are scored.
Downloads
Files actually published for this run. Original output, site-made previews and post-run conversions are labelled separately.
The exact prompt
as run, 1,074 characters. Every result for this prompt
Build a real-time 3D world in Three.js that looks like a living impressionist oil painting (one HTML file, CDN imports only, no external assets or textures; everything procedural). The whole frame must read as painted, not as a 3D render with a filter: a brush-stroke post-process (oriented, varied-size strokes that follow surface direction and color, visible canvas grain), a painterly sky with broken-color clouds, water with dappled reflections of the sky and shore, and a small landscape (garden, footbridge or shoreline, trees, flowers). Add falling snow that is painted too, and flocks of birds flying through the scene. Include two or three painted figures; clicking one opens a short in-world dialogue in a hand-lettered caption style. Interactive camera (orbit plus a gentle walk mode), a time-of-day slider that shifts the palette from cool morning to warm dusk, and a toggle that compares the painted view with the plain render. Must stay smooth on a laptop and usable on a phone (touch controls, adaptive quality). It should feel like stepping into a painting.
How it was made
Step by step
The prompt, every agent turn and tool call (tool name and a one-line summary; inputs and outputs are not shown), then the final message. The full transcript is 5.9 MB; it is published as a scrubbed, compressed download rather than shown on the page.
Show all 116 steps
Loading the timeline…
Download the full transcript OpenCode session export, 77 KB, gzip. Scrubbed: anything that looks like a key, token, e-mail address or phone number is replaced, and account details and the test machine's file paths are removed.
Cost breakdown
Headline cost $3.00, as reported by OpenCode (its own bill for the run).
| Tokens by type | Tokens | $ per 1M | Cost |
|---|---|---|---|
| Input (uncached) | 74,241 | n/a | n/a |
| Cache write | 0 | n/a | n/a |
| Cache read | 3,651,413 | n/a | n/a |
| Output | 45,671 | n/a | n/a |
- Total computed from tokens
- unknown
- Total reported by the harness
- $3.0048
OpenCode 1.18.35 reported its own total; token counts x list prices shown beside it. No list price for kimi-k3-fast in config/prices.json; no estimate computed. The harness reports reasoning tokens separately from output; here output includes them (32,474 visible + 13,197 reasoning), and both bill at the output rate.
Run notes
- frozen prompt (written): core requirements sha256 1ddb514daa65, full harness prompt sha256 1ddb514daa65; exact prompt text used by the existing harness rows of this test on the runs branch (copied from threejs_impressionist_world__claude-code__opus-5-5-high__attempt1/meta.json, sha256 verified); Beta hal
- no TUI: the harness ran in its non-interactive mode, so there are no TUI screenshots; the timelapse is built from version renders only
- OpenCode config (providers only, permission allow-all except question; no agents/instructions) + both default skill kits in xdg/config/opencode/skill (skills=on)
- infra retries: 0
- in the published transcript the sandbox HOME directory is shown as <HOME>
- skills=on: 54 skills installed in the harness skill directory before launch (video-33 remotion-dev/skills + heygen-com/hyperframes@0b5db9da + f; web-threejs-21 [redacted]/ai-AI Model Reviewer skills/ (alpha/mr-homebar)@7521074d98c9).
- Mode: headless run mode (opencode run --format json), inside bubblewrap sandbox
- Isolation audit: none
- Renderer: Three.js 0.160.0 on WebGL (loaded from a CDN) (read from the run's own 001_20261008T225156Z.html, 002_20261008T225451Z.html, 003_20261008T225856Z.html, 004_20261008T225901Z.html).
- Card video: none (the run has no H.264 yuv420p video.mp4 with 30+ frames under 24 MB). Poster: a headless-Chromium screenshot of the run's own index.html taken at import (the run folder has no render.png).
- Status: stopped by us. Stopped at the row cap (45 min) before the harness declared it finished; the page shown is what it had built by then.
- Skills on: the run had 54 skills installed (33 video + 21 web/Three.js), as its meta records.
- Kimi K3 Fast is a fast-serving router over the kimi-k3 weights (Fireworks router accounts/fireworks/routers/kimi-k3-fast; responses report kimi-k3), not separate weights
- Kimi K3 (fast router) served via Fireworks (OpenCode provider fireworks-ai).
- Same frozen prompt as the other rows on this test; configurations differ and all rows are unscored, so this is not a head-to-head result.