velaia/Kolibri-1-MLX-2bit community quant repo

Fields

id

sources/hf-velaia-kolibri-1-mlx-2bit

URL

https://huggingface.co/velaia/Kolibri-1-MLX-2bit

Publisher

velaia

Kind

repo

Published

2026-10-03

Accessed

2026-10-17

notes

Unofficial MLX conversion of Aleph-Alpha/Kolibri-1, read 2026-10-17. The card own table: 4-bit is 4.54 bits/weight, 41 GiB, peak memory 44/48 GB (short prompt / 4k-token prompt), fits 64 GB+ Macs; 3-bit is 33 GiB, peak 35/39 GB, fits 48 GB+ Macs; 2-bit is 2.61 bits/weight, 24 GiB, peak 26/30 GB, fits 36 GB+ Macs. Perplexity on a 4,096-token Wikipedia sample is reported as DE/EN 12.77/16.47 (4-bit), 13.17/16.31 (3-bit), 14.01/17.46 (2-bit). Generation about 52-56 tok/s on an M1 Max. Not affiliated with Aleph Alpha. A runnability and fidelity source, not a published-benchmark source.

Notes

Unofficial MLX conversion of Aleph-Alpha/Kolibri-1, read 2026-10-17. The card own table: 4-bit is 4.54 bits/weight, 41 GiB, peak memory 44/48 GB (short prompt / 4k-token prompt), fits 64 GB+ Macs; 3-bit is 33 GiB, peak 35/39 GB, fits 48 GB+ Macs; 2-bit is 2.61 bits/weight, 24 GiB, peak 26/30 GB, fits 36 GB+ Macs. Perplexity on a 4,096-token Wikipedia sample is reported as DE/EN 12.77/16.47 (4-bit), 13.17/16.31 (3-bit), 14.01/17.46 (2-bit). Generation about 52-56 tok/s on an M1 Max. Not affiliated with Aleph Alpha. A runnability and fidelity source, not a published-benchmark source.

Cross-references

This record references no other records.