By use case · Multilingual
Best local multilingual models
These models are trained beyond English, so they hold up in Indic, Korean, or Chinese where an English-first model drifts. 6 open multilingual models are ranked by quality below. The strongest is Sarvam-105B (~67.5 GB at Q4_K_M); the lightest is Sarvam-1 2B (~2.7 GB).
Models with language-first training beyond English (Indic, Korean, Chinese).
- 1 S ~67.5 GBSarvam-105B105B MoE · runs on 4/40 devices
- 2 EX ~20.2 GBEXAONE 4.0 32B32B · runs on 13/40 devices
- 3 S ~21.7 GBSarvam-30B30B MoE · runs on 12/40 devices
- 4 S ~16.3 GBSarvam-M 24B24B · runs on 13/40 devices
- 5 ER ~14.3 GBERNIE 4.5 21B-A3B21B MoE · runs on 17/40 devices
- 6 S ~2.7 GBSarvam-1 2B2B · runs on 40/40 devices
Tagged by each model's stated purpose. The memory figure is what it needs at Q4_K_M, and the device beside it is the lightest tracked machine that fits it. "Runs on N devices" counts the 40 tracked devices that fit it at Q4_K_M.
FAQ
What is the best local multilingual model?
Sarvam-105B leads on quality among the open multilingual models tracked here. It needs ~67.5 GB at Q4_K_M, so the lightest hardware that runs it is Apple M4 Max (128GB). Pick by what fits your memory using the list above.
What is the smallest multilingual model that runs on a laptop?
Sarvam-1 2B is the lightest, at ~2.7 GB at Q4_K_M, so any device with 8 GB or more can load it.
How were these multilingual models chosen?
They are open-weight models whose own design targets multilingual (by name, family or model card). Memory figures are computed at Q4_K_M and sourced; see the methodology page.
Memory is computed at Q4_K_M, validated 2026-08-03. See methodology.