InclusionAI released Ling-3.1-flash, a hybrid reasoning model with 560B total parameters. It activates 25B parameters per token and supports a 1M-token context window. The model improves on real-world tasks like agents, search, office work, and software engineering. It scores well on GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, and Terminal-Bench 4.0. It also performs strongly in finance and materials science. InclusionAI introduced LingCode, a harness for long-horizon coding tasks. Ling-3.1-flash works with SGLang and vLLM for inference. Verified recipes exist for H200, GB300, H800, H100, and B200 GPUs. The model is available on Hugging Face, ModelScope, and OpenRouter.
What changed
Ling-3.0-flash was the prior version with lower real-world task performance.
What it unlocks
Running complex, long-horizon coding tasks with hundreds of tool calls and repeated context compaction.
- 560B total parameters
- 25B active parameters per token
- 1M-token context window
- 63 KDA + 9 Gated MLA layers
What you need to act on it
- SGLang or vLLM deployment
- 8× H200/GB300 or equivalent GPU setup
Sources