Sarashina GGUF models
99 GGUF models in the Sarashina family, quantized by 34 uploaders. Sizes run from 793M to 70.1B parameters. The longest trained context is 16K tokens.
About the Sarashina GGUF models
Sarashina is represented on ggufmodels.com by 99 GGUF models, quantized by 34 uploaders on Hugging Face. The most downloaded are Sarashina2.2 3b Instruct v0.1, Sarashina2.2 1b Instruct v0.1 and Sarashina2.2 0.5b Instruct v0.1. Parameter counts run from 793M to 70.1B, and the longest trained context window is 16K tokens.
A Q4_K_M file of a Sarashina model is between 0.5 to 42.2 GB, which is the quant most people start with. Most files come from mradermacher (34), FiShota (8) and keisuke-miyako (7). The quants you will meet most often are Q4_K_M (89), Q5_K_M (69), Q6_K (64) and Q8_0 (59), where the number is how many models offer that quant.
Frequently asked questions
Which Sarashina GGUF should I download?
The most downloaded Sarashina GGUF here is Sarashina2.2 3b Instruct v0.1. Pick the size that fits your GPU first, then choose a quant: Q4_K_M is the usual starting point, and Q5_K_M or Q6_K give more quality if you have memory to spare. The VRAM pages below list what fits each GPU.
How much VRAM does a Sarashina model need?
Q4_K_M files of Sarashina models on this site range from 0.5 to 42.2 GB. Add the KV cache for your context length and about 0.5 GB for the runtime, which usually means 1.5 to 3 GB on top of the file at an 8K context.
Where do Sarashina GGUF files come from?
The most common quantizers here are mradermacher, FiShota and keisuke-miyako. They convert the original model and publish quantized files on Hugging Face. ggufmodels.com does not host files; every download button opens Hugging Face.