XHToken-Spark-X2.6-1.7B-GGUF

Spark-X2.6-1.7B (XHToken/Spark-X2.6-1.7B) is an open-weight 1.7B-parameter language model from the SparkLLM Team, released under Apache-2.0 for efficient on-device deployment across English, Chinese, and multilingual workloads. Built on Spark-X2.6-1.7B-Base, it retains the sparkx2_5 architecture of Spark-X2.5-1.7B while introducing an upgraded post-training pipeline for coding, software engineering, mathematics, general knowledge, complex tool use, agentic planning, and long-context reasoning. It supports context windows up to 1,048,576 tokens and 131,072 output tokens, with chain-of-thought reasoning enabled by default and configurable thinking levels (none, low, and high) to balance quality, latency, and cost. In edge evaluations, Spark-X2.6-1.7B delivers higher decoding throughput than comparable models, including Qwen3.5-2B, MiniCPM-2B, and LFM2.5-2.6B, with further acceleration from XHToken's optimized llama.cpp fork on an NVIDIA RTX 4090. The model supports Hugging Face Transformers (transformers==4.57.1), SGLang, vLLM with the Spark plugin, Huawei Ascend NPUs, Spark-MLX-LLM, llama.cpp (b10828+), Ollama (v0.34.1+), LM Studio (0.4.0+), and Unsloth Studio. Fine-tuning is supported through XHToken's LLaMA-Factory fork.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp