---
base_model: mistralai/Mistral-Nemo-Instruct-2407
library_name: transformers
quantized_by: InferenceIllusionist
language:
  - en
  - fr
  - de
  - es
  - it
  - pt
  - ru
  - zh
  - ja
tags:
- iMat
- gguf
- Mistral
license: apache-2.0
---
<img src="https://i.imgur.com/P68dXux.png" width="400"/>

# Mistral-Nemo-Instruct-12B-iMat-GGUF

<b>Important Note: Inferencing is *only* available on this fork of llama.cpp at the moment: https://github.com/iamlemec/llama.cpp/tree/mistral-nemo (All credits to iamlemec for his work on Mistral-Nemo support)

Other front-ends like the main branch of llama.cpp, kobold.cpp, and text-generation-web-ui may not work as intended</b>

Quantized from Mistral-Nemo-Instruct-2407 fp16.
* Weighted quantizations were creating using fp16 GGUF and groups_merged.txt in 92 chunks and n_ctx=512
* Static fp16 will also be included in repo

For a brief rundown of iMatrix quant performance please see this [PR](https://github.com/ggerganov/llama.cpp/pull/5747)

<i>All quants are verified working prior to uploading to repo for your safety and convenience</i>

<b>KL-Divergence Reference Chart</b>
 (Click on image to view in full size)
[<img src="https://i.imgur.com/mV0nYdA.png" width="920"/>](https://i.imgur.com/mV0nYdA.png)


<b>Tip:</b> If you are getting a `cudaMalloc failed: out of memory` error, try passing an argument for lower context in llama.cpp, e.g. for 8k: `-c 8192`

If you have all ampere generation or newer cards, you can use flash attention like so: `-fa`

Provided Flash Attention is enabled you can also use quantized cache to save on VRAM e.g. for 8-bit: `-ctk q8_0 -ctv q8_0`


Original model card can be found [here](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407)