Model history
Space Bunny Evals: Judging a Model With No Vendor
A model with no name, no price and no benchmarks got judged entirely on what it built — and a careful re-test showed the method was real, not memorised.
Why the anonymous window was a free evaluation
For 25.8 days, MiniMax M3.1 Flash ran on OpenRouter under the name Space Bunny Alpha with no vendor, no price, no benchmark table and no documentation. That sounds like a weakness. It was also the most honest product launch a model company can run: the model was judged only on what it actually did, in public, by strangers who had no incentive to be kind about it.
The community’s verdict came down to a handful of demos that spread on X, and then — after MiniMax claimed the model — a deliberately adversarial re-test.
Everything on this page is third-party measurement. We did not run these tests ourselves, and we have no reproduction. The main evaluation is published by 光子星球 (Photon Planet), the research brand of 钛媒体 (TMTPost); the demo links and reactions are from the Hacker News launch thread.
The five demos that made the model famous
These are the anonymous-era projects the community circulated. They are worth knowing individually, because the re-test later swaps each one for a new task of the same type.
| Demo | What it was | What the community made of it |
|---|---|---|
| Rocket launch | A cinematic 3D scene, claimed to be built in about three minutes, and compared against Opus 5 on the same prompt | The flagship comparison of the whole episode |
| Voxel garden | A voxel-art garden scene, watched by roughly 100,000 people | Looked good at a glance; on inspection the peach trees, water and pavilion detailing were rough |
| GPU fluid simulation | A single HTML file, zero dependencies, 1,564 lines of hand-written code, simulating GPU-style fluid | The most widely circulated artefact of the anonymous period |
| Mechanical bee | A precision mechanical bee with interlocked parts and realistic materials | The only case the community judged better than a flagship model, on mechanical precision and materials |
| 3D racing game | “Build me a 3D racing game”, one sentence | The famous failure: it ran, but clipped through geometry and steered backwards; one blogger iterated three times without fixing it |
Notice the shape of that table. Three hits, one qualified hit, and one clear failure. This is what a real evaluation looks like — it is not a demo reel.
The re-test: swapping the question to break the memorisation
The obvious objection to the demos is that they had circulated for a month. The model had very likely seen the original prompts. Re-running them would measure recall, not capability.
So the evaluation changed the questions and kept the structure. Same task types, same difficulty, same judging dimensions — different subject matter:
| Original demo | Re-test task | What the swap tests |
|---|---|---|
| Rocket launch | Volcanic lava flowing into the sea | Harder: the lava must interact continuously with the water and the steam must keep emitting rather than bursting once |
| Voxel garden | A Jiangnan water town | Adds a water-reflection layer, removes the large plant areas, and demands coherent building structure and material layering |
| GPU fluid | A particle galaxy | Whether the model grasps the general method of graphics programming or had memorised one fluid implementation |
| Mechanical bee | A mechanical pocket watch | The bee tested “does it look like”; the watch tests “is it correct” |
| 3D racing game | Kept as the original | It was the community’s known failure, so the point was to see whether the old bug would come back |
That last choice is the sharpest one in the whole design. A benchmark that only runs the successes proves nothing; a benchmark that deliberately re-runs the failure is the one that carries information.
The results
Volcanic lava into the sea
The first version ran. The camera came in low over the water, rose across the lava river, and pulled back to frame the volcano, the coast and the steam — the three-stage shot was delivered. Steam emitted continuously from the contact point and was slanted by the sea wind, denser at the bottom and thinner at the top rather than a single white column. Sparks arced up and faded out. The sea reflected the orange glow.
Where it fell short was the lava itself. The prompt asked for the lava river to split into three streams at its mouth. What came back was closer to three glowing bonfires sitting there, with no sense of flow. The surface crust and glowing cracks were good; the volume rendering and the steam were the hard bones and it handled them. The morphology was the miss.
The Jiangnan water town
This one came back near full marks, and the published rating was 4.5 out of 5. The details that earned it: a stone arch bridge with actual steps, stepped horse-head walls, white walls with black tiled sloped roofs, two awning boats drifting slowly enough to read the movement over ten seconds, willow branches trailing down to the water, red lanterns hanging under the eaves, stone steps running down to the water’s edge.
The day/night toggle lit the lanterns and windows and warmed the reflections with a smooth gradient rather than a hard cut. The water rippled. The willows swayed. Even the angle of the sunlight changed between day and night.
The original voxel garden had the same kind of rough edges on close inspection. It did not have them here.
The particle galaxy
This is the one that most cleanly separated “method” from “answer”. The fluid demo and the galaxy demo are not the same problem: one is grid-based advection, the other is position-based state update.
The result: 100,000 particles in a flat spiral arm rotating slowly; dragging the mouse creates a gravity source that pulls nearby particles into a vortex; releasing lets them drift back to orbit gradually rather than snapping back. Velocity colouring was correct — slow particles blue, fast ones white-yellow, fast swings leaving a short trail without smearing.
And the part that matters most for real work: there were no dropped frames. Particle-effect web projects are exactly where this kind of build usually falls apart on performance, and it did not.
The mechanical pocket watch
The bee asked “does it look like”. The watch asked “is it right”. The prompt required more than twenty interlocking gears, an oscillating balance wheel, and working hands — and specified that a single wrong element would kill the whole thing.
What came back: real mechanical structure, gears genuinely interlocked with adjacent gears counter-rotating, the balance wheel oscillating, the escape wheel and pallet fork present, and hour, minute and second hands all running correctly with the second hand ticking once per second. The case opens on a hinge, showing polished metal reflection on the inside and worn patina on the outside, Roman numerals correctly placed, and the whole thing orbits when you drag it.
The published verdict: a high-completion project. And the community’s claim that the mechanical bee beat a flagship model “is not an exaggeration” when applied to the watch.
The racing game
The one original case that was deliberately kept, because it was the community’s failure.
It did not fail this time. The published account says the model set its own standard at the start of the task, and that during the fixing phase it actively opened a browser to test its own work. The self-checking behaviour the community had praised during the anonymous period survived into the official release.
The flaw that remained: the handbrake was hard to use, with weak feedback — and, crucially, that kind of detail was still not caught by its own self-review. The bug from the original case being fixed is a stronger signal than luck.
What the re-test actually proved
Three conclusions, and one warning.
- It is a method, not an answer. Four new tasks of the same type, and the model still delivered. Memorised answers expire; a method does not. This is the single most useful thing a re-test can establish.
- Self-verification is the most valuable behaviour. Opening a browser to test its own changes puts the model a tier above “edit the code and submit”. It is the difference between a demo and a deliverable.
- The cost case holds up. A model that can do front-end work at this level, at M3.1 Flash’s price, is a genuinely capable option for medium-difficulty daily coding work.
- But nobody had a vendor to ask. For three and a half weeks the model was excellent at this and completely unattributable. The evaluation worked; the accountability did not exist. That is the part worth remembering.
What it does not prove
- No reproduction. These are one publication’s results, on a model that has since changed. We have not re-run any of it, and neither has anyone else we can point to.
- Front-end is not the whole model. Every test here was front-end graphics and interaction. Nothing in this evaluation speaks to long-document reasoning, tool-heavy agent loops, or the coding work the benchmarks page covers.
- The 3D demos are not a benchmark. A prompt that produces an impressive scene says something about the model’s ceiling on a good prompt. It says nothing about the median case, the error rate, or what happens on the twentieth request in a session.
- A model can be good at the showcase and ordinary at the work. That gap is exactly what the M3.1 Flash benchmarks and the use cases page are for.
Where this leaves M3.1 Flash
MiniMax launched M3.1 Flash on 19 October 2026 with the claim that it makes online inference nearly three times faster than its predecessor and cuts agentic task cost by 70%, at $0.15 per million input tokens, $0.60 output and $0.015 cached. The official scores are BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, Agents’ Last Exam 29.6 and CyberGym 82%.
The anonymous window is the strange part of that launch: the model was already being judged on its output for nearly a month before anyone knew whose it was. If you want to know what the community thought of it while it was still nameless, start with the five demos, then read the Space Bunny story for how it was identified, and use the benchmarks page for what has been independently measured since.
Sources and version notes
Checked October 21, 2026. We label official claims, third-party measurements, our own testing and community reports separately — see the sources policy for what each label means.