EXAONE 3.5 2.4B Instruct — GGUF + MXQ for llama-cli-mblt

This repository provides EXAONE 3.5 2.4B Instruct compiled and optimized for Mobilint NPU hardware, packaged for use with llama.cpp-mblt.

Branches

Branch	Contents	Description
`main`	Body model only	Standard autoregressive decoding
`eagle3`	Body + FC + Draft models	EAGLE3 speculative decoding (~2-4x faster)

Quick Start

# Simple decoding
llama-cli-mblt -hf mobilint/EXAONE-3.5-2.4B-Instruct-GGUF -p "Hello!" -n 128

# EAGLE3 speculative decoding
llama-cli-mblt -hf mobilint/EXAONE-3.5-2.4B-Instruct-GGUF --eagle3 -p "Hello!" -n 128

# Interactive chat
llama-cli-mblt -hf mobilint/EXAONE-3.5-2.4B-Instruct-GGUF --eagle3

Files

main branch

File	Size	Description
`exaone-3.5-2.4b-instruct-vocab.gguf`	4.0 MB	Tokenizer (vocab-only GGUF)
`target_emb.bin`	1.0 GB	Body embedding weights (float32)
`EXAONE-3.5-2.4B-Instruct.mxq`	1.4 GB	Body model for NPU
`config.json`	—	Model configuration

eagle3 branch (adds)

File	Size	Description
`single_Fc_EXAONE-3.5-2.4B-Instruct.mxq`	19 MB	FC dimension converter model
`Draft_EXAONE-3.5-2.4B-Instruct.mxq`	87 MB	EAGLE3 draft model
`draft_emb.bin`	1.0 GB	Draft embedding weights
`d2t.bin`	250 KB	Draft-to-target vocabulary mapping