miiqu-gguf

miiqu-105b-v1.0

Developed by Infinimol AI GmbH

Also Available:

• EXL2: 5.0bpw
• F16: HF

8th place on EQ-Bench, beating Qwen1.5-72B-Chat, miqudev/miqu-1-70b, mistral-medium and claude-3-sonnet-20240229. All without fine-tuning or additional training.

Thanks for support from: turboderp, silphendio, sqrkl, and ngxson!

❗ Q4_K_M files are split and require joining

Note: HF does not support uploading files larger than 50GB. The Q4_K_M files are supplied as split files.

<details><summary>Click for instructions regarding Q4_K_M files</summary>

Process

Please download:

• miiqu.gguf-split-aa
• miiqu.gguf-split-ab
• miiqu.gguf-split-ac
• miiqu.gguf-split-ad
• miiqu.gguf-split-ae
• miiqu.gguf-split-af

To join the files, do the following:

Linux and macOS:

cat miiqu.gguf-split-a* > miiqu_Q4_K_M.gguf && rm miiqu.gguf-split-a*

Windows command line:

COPY /B miiqu.gguf-split-aa + miiqu.gguf-split-ab + miiqu.gguf-split-ac + miiqu.gguf-split-ad + miiqu.gguf-split-ae + miiqu.gguf-split-af miiqu_Q4_K_M.gguf
DEL miiqu.gguf-split-aa miiqu.gguf-split-ab miiqu.gguf-split-ac miiqu.gguf-split-ad miiqu.gguf-split-ae miiqu.gguf-split-af

</details>

Model Details

• Max Context: 32768 tokens
• Layers: 105

Prompt template: ChatML or Mistral

chatml:

<|im_start|><|user|>\n<|user-message|><|im_end|>\n<|im_start|><|bot|>\n<|bot-message|><|im_end|>\n

mistral:

[INST] <|user|><|user-message|>[/INST]<|bot|><|bot-message|></s>