katalidevai/spark-x2.5-4b-katali-q4

Spark-X2.5-4B — Katali EQS (q4)

A quantized, Katali-format build of XHToken/Spark-X2.5-4B, a dense on-device model with native long-context support, for local chat with no GPU/VRAM required.

⚠️ Format note

These are EQS files — Katali's own packed model format, not safetensors or GGUF. They only work with the Katali app; they are not compatible with llama.cpp, Ollama, vLLM, or other runtimes.

What was changed from the original

Dense weights (attention + MLP) were quantized to q8_0 and packed into Katali's EQS shard format. No weights were retrained or fine-tuned — this is a lossy quantization of the original released weights, not a derivative model in the training sense.

Files

File	Description
dense.eqs	Quantized dense weights
config.json	Model architecture config
tokenizer.json, vocab.json, merges.txt, tokenizer_config.json, special_tokens_map.json	Tokenizer
chat_template.jinja	Chat template
generation_config.json	Default generation settings

How to use

1. Download this repo into a folder, e.g. Models\spark-x2.5-4b-katali-q4\
2. In Katali, Browse to that folder and press Load (Device: Host)

License

Apache License 2.0, same as the original XHToken/Spark-X2.5-4B. See LICENSE for the full text. All credit for the base model goes to its original authors.