Glass aquarium burst — cracked panel, hydrostatic jet
One line of prompt, four browser sims of the same accident: a side panel cracks, the glass gives, and the tank empties. The point of the post is the top-left card — Kimi K3 quantised to 1 bit, running on the author's own hardware, against two frontier APIs. Reasoning effort was not held equal across the four, so read the clips as four attempts at one scene, not as a race.
The prompt behind "Glass aquarium burst — cracked panel, hydrostatic jet" is published in full, and was run by 4 models: 1-bit Kimi K3, Kimi K3, GPT-5.6 and Opus 5. 4 of 4 finished, on figures reported by the author.
- Models
- 4
- Completed
- 4 of 4
- Cheapest completed run
- not measured
Figures reported by Unsloth AI. Not independently verified.
Numbers on this page were reported by the author. We did not run this comparison and cannot verify them.
We did not run this. The prompt line above is the one the author published verbatim in the post; no hash is recorded because the exact string sent to each model is not public. No cost, token count or duration was published for any of the four runs, so all three are left unmeasured here rather than reconstructed from list prices. The one figure the author does give is a throughput, not a price: the 1-bit Kimi K3 GGUF (UD-IQ1_M, 594 GB, which the companion post puts at ~78.9% of the full model's accuracy after shrinking it from 1.56 TB) ran locally on 4x B200s at 36 tok/s. Reasoning effort differs between the runs and is burnt into the source video: MAX for both Kimi K3 runs, EXTRA HIGH for GPT-5.6, HIGH for Claude Opus 5 — the four are not the same setting, which is why nothing here is aggregated or ranked. The source is a single 2x2 video; each clip is that video's own quadrant for one model, cropped, over the same 1.5-9.5s window so the four show the same moment of the same accident. The last third of the source video is a screen recording of Unsloth's own app and is not published here.
Output
The prompt that built Glass aquarium burst
Create a glass aquarium whose side panel develops a visible crack and then bursts.
Runs
| Model | Harness | Status | Cost | Duration | Output | Trust | Artifact |
|---|---|---|---|---|---|---|---|
| 1-bit Kimi K3kimi-k3-1bit | — | Completed | not measured | not measured | not measured | Reported | — |
| Kimi K3kimi-k3 | — | Completed | not measured | not measured | not measured | Reported | — |
| GPT-5.6gpt-5-6 | — | Completed | not measured | not measured | not measured | Reported | — |
| Opus 5claude-opus-5 | — | Completed | not measured | not measured | not measured | Reported | — |