MI

Mistral 22B v0.2

MistralGGUF by failspyUpdated 2024-04-13
Download on Hugging Face
Parameters22.2B
Context64K
Layers56
Architecturellama
Smallest quant8.3 GB
Quick answer

Mistral 22B v0.2 needs about 15.7 GB of memory at Q4_K (a 13.3 GB file) with an 8K context window, so it fits on a 16 GB GPU. This repository has 11 quants, from 8.3 GB (Q2_K) to 44.5 GB (F16). Longer context needs more: at 32K tokens the KV cache alone adds about 7.5 GB.

Mistral 22B v0.2 is a 22.2B-parameter Mistral model in GGUF format, quantized by failspy. Its trained context length is 64K tokens. The repository has 11 quant files, from 8.3 GB (Q2_K) to 44.5 GB (F16).

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q2_K8.3 GB 10.6 GB
Too big Get file →
Q3_K_S9.6 GB 12.0 GB
Too big Get file →
Q3_K10.8 GB 13.1 GB
Too big Get file →
Q3_K_L11.7 GB 14.1 GB
Too big Get file →
Q4_K_S12.7 GB 15.0 GB
Too big Get file →
Q4_K13.3 GB 15.7 GB
Too big Get file →
Q5_K_S15.3 GB 17.7 GB
Too big Get file →
Q5_K15.7 GB 18.1 GB
Too big Get file →
Q6_K18.2 GB 20.6 GB
Too big Get file →
Q8_023.6 GB 26.0 GB
Too big Get file →
F1644.5 GB 46.9 GB
Too big Get file →

What GPU do you need for Mistral 22B v0.2?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: the best quant that fits is Q3_K_S (9.6 GB file, about 12.0 GB in use).
  • 16 GB VRAM: the best quant that fits is Q4_K (13.3 GB file, about 15.7 GB in use).
  • 24 GB VRAM: the best quant that fits is Q6_K (18.2 GB file, about 20.6 GB in use).
  • 32 GB VRAM: the best quant that fits is Q8_0 (23.6 GB file, about 26.0 GB in use).
  • 48 GB VRAM: the best quant that fits is F16 (44.5 GB file, about 46.9 GB in use).

Which Mistral 22B v0.2 quant should you choose?

Q4_K (13.3 GB) is the middle option in this repository. The usual sweet spot between file size and quality, and the quant most people start with.

If memory is tight, the smallest file is Q2_K at 8.3 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturellama
Parameters22.2B
Layers56
Trained context65,536 tokens
KV cache (fp16)224 KB per token

Source

Quantized byfailspy
FamilyMistral
Licenseapache-2.0
Downloads (HF)182
Last updated2024-04-13

Frequently asked questions about Mistral 22B v0.2

How much VRAM does Mistral 22B v0.2 need?

About 15.7 GB at Q4_K with an 8K context window. That is the 13.3 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Mistral 22B v0.2 on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 10.6 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Mistral 22B v0.2 GGUF?

It depends on the quant: Q2_K is 8.3 GB, Q3_K_S is 9.6 GB, Q3_K is 10.8 GB, Q3_K_L is 11.7 GB, Q4_K_S is 12.7 GB, Q4_K is 13.3 GB and 5 more.

What context length does Mistral 22B v0.2 support?

The GGUF header lists a trained context length of 65,536 tokens (64K). The KV cache grows with the context you use: about 1.9 GB at 8K and 7.5 GB at 32K.

Where can I download Mistral 22B v0.2 GGUF?

The files are on Hugging Face in the failspy/Mistral-22B-v0.2-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Mistral 22B v0.2 use?

Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Similar models

Browse all →

More from failspy

All by failspy →

More Mistral models

All Mistral →

Other medium models

All medium models →