Spark-X2.6-4B brings stronger on-device agents and coding

Spark-X2.6-4B is out: a compact general-purpose model for on-device agents with post-training gains in coding, tool use and long-context tasks. It scores 62.8 vs 41.6 on SWE Verified and 71.8 vs 65.1 on BFCL-V4 over Spark-X2.5-4B. The model is Apache 2.0 and available on Hugging Face, ModelScope and Ollama.

• SWE Verified: 62.8 vs 41.6 for Spark-X2.5-4B
• BFCL-V4: 71.8 vs 65.1 for the previous version
• Three thinking levels: none, low and high
• Apache 2.0 license, runs via Ollama and llama.cpp