Compact 1.7B open-weights text LLM for on-device use, with adjustable thinking levels and English/Chinese support.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Spark-X2.6-1.7B is a small general-purpose language model published on Hugging Face by XHToken (Spark LLM project). It is the second model in the Spark-X2.6 series, following Spark-X2.6-4B. It keeps the same architecture as Spark-X2.5-1.7B, and the changes come from improved post-training, aimed at coding, agentic tool use, long-context, multilingual and reasoning tasks.
The model card describes three configurable thinking levels (none, low, high) so applications can trade reasoning quality against latency and cost. According to the card, switching from high to low reasoning effort lowers the average benchmark score by 4.0 points and reduces average output length by 37.3%. The vendor reports leading decoding throughput among similarly sized models (Qwen3.5-2B, MiniCPM-2B, LFM2.5-2.6B) and provides an optimized llama.cpp fork. These benchmark and speed claims are vendor-reported. A YouTube creator described it as about 3.4 GB and fast on a Mac M5 Pro and an RTX 4090.
Weights are distributed in safetensors format under Apache-2.0 and load through Hugging Face Transformers (the card specifies version 4.57.1, with custom code). Tags list English and Chinese. It is a text-only model aimed at developers running local or edge inference. A base model, Spark-X2.6-1.7B-Base, is also referenced.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
1.7B on-device model that beats similarly sized models such as Qwen 3.5 2B and Gemma 4 on many benchmarks and is fast on Mac M5 Pro and RTX 4090. About 3.4GB.
Closest neighbours by meaning, not just by category.
Model · OpenBMB
Dense 2B-parameter open-weights language model from OpenBMB for on-device and edge deployment, Apache-2.0 licensed.
Model · Alibaba
Small 0.8B-parameter open-weights language model in Alibaba's Qwen 3.5 family, aimed at lightweight and on-device use.
Model · OpenBMB
1B-parameter dense open-weights language model from OpenBMB for on-device use, with hybrid thinking and 128K context.
Model · Liquid AI
Liquid AI's 24B-parameter mixture-of-experts LFM with 2B active parameters, built for fast on-device agents.
Model · Meta
Meta's open-weight 816B-parameter mixture-of-experts model (42B active, 1M context) in the Muse model family.
Model · OpenBMB
Small open-weights multimodal model from OpenBMB for image and video understanding on phones and edge devices.