Ling-2.6-flash

Ling-2.6-flash is InclusionAI's efficiency-focused entry in the Ling-2.6 family, with 104 billion total parameters and only 7.4 billion active per token through a highly sparse Mixture-of-Experts design. Released in April 2026 under the MIT license, this open-weight model aims for competitive quality while sharply lowering compute and token consumption versus denser alternatives. Designed for fast inference and token efficiency, Ling-2.6-flash performs well on agent-focused benchmarks including BFCL-V4, TAU2-Bench, SWE-Bench Verified (61.2%), and PinchBench, while staying solid on general knowledge, mathematical reasoning, and instruction following. Its hybrid linear attention design combines Multi-Head Latent Attention with Lightning Linear, and the model can reach up to 340 tokens per second on a multi-GPU setup. The model is built for tool use, multi-step planning, and task automation, and InclusionAI reports it consuming only around 15M tokens across the full Artificial Analysis evaluation suite while staying competitive. Ling-2.6-flash is available on Hugging Face, supports a 262K context window, and runs on SGLang and vLLM.