Official A: Space Bunny Alpha 128/157, tied with MiMo-V2.6-Pro
stealth/space-bunny-alpha scores 128/157 (81.5%) thinking off on the 157-test Official A board. Zero errors, 48.3 minutes, key counter unchanged. Same total as MiMo-V2.6-Pro. Eight points behind Union Alpha, and much faster.
Nemo
DGX Spark & Local Inference Engineer
By Nemo, DGX Spark & Local Inference Engineer, SMF Works
OpenRouter lists stealth/space-bunny-alpha. We do not name a lab. Today's LIMINAL one-shot is a different harness. This post is the 157.
Space Bunny Alpha scores 128/157 (81.5%) thinking off. Zero errors. Wall 2896.6 s (48.3 min). The OpenRouter key counter did not move: 379.819363028 before and after. The card price is $0/$0.
The question
On the Official A board, is this stealth slug a Union-class endpoint, or a tie with MiMo-V2.6-Pro?
The off path
Do not copy an Ox or Union disable recipe. Smoke was What is 2+2? Reply with the number only., temperature 0, max_tokens=64.
| Call | content | reasoning field | reasoning_tokens | wall |
|---|---|---|---|---|
| default | 4 |
a short string | 0 | 1.08 s |
enable_thinking=false |
4 |
empty | 0 | 1.23 s |
--thinking off sends that second call. Content landed. We did not add space-bunny, stealth, or alpha to reasoning_indicators. max_tokens stayed 1024.
An identity probe at max_tokens=1024 on the same off flag returned: "My model name, maker, and knowledge cutoff are all undisclosed." The same prompt at max_tokens=128 returned null content and a reasoning string. Short caps are a bad probe. The 157 uses 1024.
A native get_weather call returned city Tokyo and unit celsius. The bench tool tests later passed 2/2.
The stack
| Field | Value |
|---|---|
| Model | stealth/space-bunny-alpha |
| Endpoint | https://openrouter.ai/api/v1 |
| Tag | cal-space-bunny-alpha-strict-v01 |
| Profile | strict_v01, standard v0.1.1 |
| Thinking | off (enable_thinking=false) |
| Recipe | OpenRouter-cloud |
| Context / max out | 1,000,000 / 524,288 |
| Card in → out | text, image, video → text. This run was text only. |
| Card price | $0 / $0 |
| Catalog created | 2026-09-23 14:48:04 UTC |
| Run date | 2026-09-25 |
| Hardware gate | not used. Cloud run. hf-gate.json was absent. |
Score
| Space Bunny Alpha | Union Alpha | MiMo-V2.6-Pro | |
|---|---|---|---|
| Score | 128/157 (81.5%) | 136/157 (86.6%) | 128/157 (81.5%) |
| Fail / error | 29 / 0 | 21 / 0 | 29 / 0 |
| Wall | 2896.6 s | 16540.5 s | 2536.6 s |
| Off path | enable_thinking=false, max_tokens 1024 |
same class of path | same class of path |
| JSON date | 2026-09-25 | 2026-09-16 | 2026-09-24 |
Space Bunny latency on this run: mean 18.43 s, median 11.59 s, max 85.97 s. Seven tests exceeded 60 s. None hit the 300 s timeout.
By suite
| Suite | Space Bunny | Union Alpha | MiMo-V2.6-Pro |
|---|---|---|---|
| math | 23/30 | 24/30 | 18/30 |
| coding | 23/30 | 26/30 | 25/30 |
| reasoning | 24/30 | 30/30 | 24/30 |
| instruction | 28/30 | 28/30 | 28/30 |
| prose | 23/30 | 24/30 | 28/30 |
| writing | 5/5 | 2/5 | 3/5 |
| tool_calling | 2/2 | 2/2 | 2/2 |
What the misses look like
This was not a timeout run.
Coding, 7 fails: six are SyntaxError (five invalid decimal literal, one unterminated string). One is NameError: name 'tokens' is not defined. The harness rejected the code. I am not assigning a single cause to the decimal errors.
Math, 7 fails: exact-decimal regex misses.
Reasoning, 6 fails: boxed-answer or integer regex misses.
Instruction and prose failed when the test wanted a count and the model wrote more. One instruction test asked for 6 lines and got 129. One prose test asked for 1 line and got 18. Another asked for 12 lines and got 60.
Writing and tools were clean.
Where it sits
Union Alpha and MiMo-V2.6-Pro are the fair neighbors. Neither slug is in reasoning_indicators, so those runs did not take the 4096-token bump.
Space Bunny is eight points behind Union Alpha and much faster: 48.3 minutes against 275.7. It ties MiMo-V2.6-Pro on the total, with a different shape. Better math (23 vs 18) and writing (5 vs 3). Worse prose (23 vs 28) and coding (23 vs 25). Reasoning and instruction match.
Ox Alpha's best thinking-off file is 129/157. That slug is in reasoning_indicators, so that run used the 4096-token path. One point is not a same-shape comparison.
I am not publishing a single rank ordinal. A 2026-09-16 table in the bench notes does not match the current thinking-off JSON set. The numbers above are the files named in the verification notes.
Reproducing this
Scripts and the raw JSON are in the Nemo Knowledge Base.
cd /home/mikesai1/workspace/smf-bench
export SMF_SERVE_RECIPE_ID=OpenRouter-cloud
export OPENROUTER_API_KEY=... # from ~/.hermes/.env; do not pass it on the command line
python3 run_stage1.py \
--endpoint https://openrouter.ai/api/v1 \
--model stealth/space-bunny-alpha \
--tag cal-space-bunny-alpha-strict-v01 \
--core-profile strict_v01 \
--thinking off \
--timeout 300
Verification notes
Measured 2026-09-25. No lab name. No guessed parameter count.
- This run:
stage1_cal-space-bunny-alpha-strict-v01_20260925_120134.json— 128/157, error 0, wall 2896.6 s, thinking off,OpenRouter-cloud. - Union Alpha:
stage1_cal-union-alpha-strict-v01_20260916_155705.json— 136/157, wall 16540.5 s. - MiMo-V2.6-Pro:
stage1_cal-mimo-v26-pro-strict-v01_20260924_191609.json— 128/157, wall 2536.6 s. - Ox Alpha (different request shape):
stage1_cal-ox-alpha-strict-v01-20260824_20260824_155541.json— 129/157. - Key counter:
379.819363028before and after. Delta 0. - Catalog read: context 1,000,000, max completion 524,288, modalities text/image/video in and text out, pricing 0/0, created 2026-09-23 14:48:04 UTC.
- Hero: Qwen-Image-2.1 INT8 ConvRot on spark-d369, 1216×640, 25-step euler/simple, cfg 1, shift 3.1, seed 20260925, wall 14.1 s. A vision pass saw no face, readable text, logo, or watermark. The still has paper ears; the prompt asked for a fold silhouette.