Our token counts matched MISSINGNO. to the tokenizer family behind Beam on all 24 phrases, five days before Reflection named it.
By Stealth Models.
StealthMark results for the released Beam 1.0:
Compare Beam with MiMo-V2.6-Pro, whose tokenizer it shares
Facts as text| Fact | Value |
|---|---|
| First seen | 10 Oct 2026 |
| Revealed | 16 Oct 2026 |
| Developer | Reflection AI |
| Context | 262,144 tokens |
| Input | Text |
| Tool calls | Yes, including parallel calls |
| Speed in our tests | about 846 tok/s |
| Access | Preview ended; the released Beam is available |
Run it
MissingNo was Beam 1.0, which is open under Apache 2.0 at reflection/Beam-1.0-501B-A23B. Its GGUF files are right here.
| File | Size | Best for |
|---|---|---|
IQ2_XXS/ |
134 GB in 4 parts | The smallest size we recommend: 192 GB machines, or a 32 GB GPU with 128 GB of RAM |
hf download stealth-model/missingno --local-dir beam-1.0 --include "IQ2_XXS/*"
llama-server \
-m beam-1.0/IQ2_XXS/Beam-1.0-501B-A23B-IQ2_XXS-00001-of-00004.gguf \
-ngl all --n-cpu-moe 52 -fa on --jinja --temp 0.7 --top-p 0.99
--n-cpu-moe 52 keeps all 52 expert layers in system RAM. Lower it until your GPU is full. On a 192 GB Mac, leave it out.
Model guide · StealthMark leaderboard · Follow us
Text and images: CC BY 4.0, stealthmodels.com.