---
license: apache-2.0
base_model: XiaomiMiMo/MiMo-V2.5
tags:
- moe
- int4
- zstd
- peng
- colibri
---

# MiMo-V2.5 311B — int4 + lossless zstd container for peng

**109 GB instead of 152 GB (71.8%), bit-exact same model** as
[fivetech/MiMo-V2.5-colibri-peng-int4](https://huggingface.co/fivetech/MiMo-V2.5-colibri-peng-int4).

Every tensor in the safetensors shards is stored as one zstd-1 frame
(`__metadata__.peng_zstd="1"`, per-tensor `"nb"` = uncompressed size). The
[peng](https://github.com/fivetechsoft/peng-mimo) engine (commit `6ffb8d9`+) reads
this format natively and produces **bit-exact** output vs the uncompressed
container (verified: per-frame byte-compare of all 17 shards + identical TEMP=0
generations at full scale).

## Which container should you download?

| your setup | use |
|---|---|
| limited bandwidth / storage | **this one** (26% smaller) |
| plenty of RAM (experts pinned once at boot) | **this one** — decompress cost is paid once at load |
| NVMe-streaming on a core-limited box, max tok/s | the [plain int4 container](https://huggingface.co/fivetech/MiMo-V2.5-colibri-peng-int4) — measured ~18% faster streaming on a 16-core WSL2 box, because zstd decompression competes with expert matmul for cores |

Honest measurements, including the failed speed experiments:
[`findings.md` §41](https://github.com/fivetechsoft/peng-mimo/blob/peng/findings.md).

## Run

```bash
git clone https://github.com/fivetechsoft/peng-mimo && cd peng-mimo/c && make CUDA=1 mimo
SNAP=/path/to/this/repo COLI_CUDA=1 CUDA_DENSE=1 DIRECT=1 TAO=1 \
  PROMPT='Write one short sentence about Rome.' NGEN=24 ./mimo 64 4 8
```

To convert back to the plain int4 layout (or repack your own container):
`c/tools/repack_zstd.py` (`--verify` byte-compares every frame).

Derived from [colibri](https://github.com/JustVugg/colibri) (Apache 2.0, © JustVugg).