SparkLLM ships Spark-X2.6 on-device models in 1.7B and 4B
TL;DR
- SparkLLM released Spark-X2.6 on October 21 as a Hugging Face collection with 1.7B and 4B models under Apache 2.0.
- The 4B variant advertises a 1,048,576-token context window and three selectable 'thinking budget' levels: none, low, and high.
- The team's model card claims the 4B matches or exceeds ~9B models on several benchmarks, with no scores published on the collection page.
SparkLLM pushed Spark-X2.6 live on October 21, a Hugging Face collection of twelve model variants spanning 1.7B and 4B parameter sizes under Apache 2.0. Each size ships in base, instruct, GGUF, INT4, INT8 and FP8 builds, with Transformers, vLLM, SGLang, MLX, llama.cpp, Ollama, LM Studio and Unsloth support called out on the model card.
Per the model card, Spark-X2.6-4B 'achieves best-in-class performance among similarly-sized models' and 'matches or exceeds ~9B parameter models on several benchmarks', while keeping the same architecture as Spark-X2.5-4B and relying on post-training for the gains. The 4B advertises context up to 1,048,576 tokens and a 'controllable thinking budget' with three levels (none, low, high) for trading reasoning quality against latency and inference cost. The team benchmarks inference efficiency against Qwen3.5-4B and Gemma-E4B, but the collection page itself publishes no scores behind the 9B comparison claim.
Another Apache-licensed small model lining up for the on-device slot. Two researchers in our Who's Who tracker also posted the link.
Shared on Bluesky by 2 AI experts
Originally reported by huggingface.co
Read the original article →Original headline: Spark-X2.6 - a XHToken Collection