Ling-3.1-flash
inclusionAI's MIT-licensed 560B hybrid-linear MoE with 25B active parameters and a 1M-token context window.
Overview
Ling-3.1-flash is inclusionAI's hybrid-reasoning mixture-of-experts model with 560B total and 25B active parameters per token and a 1M-token context window, released on Hugging Face under the MIT license. It builds on Ling-3.0-flash with a hybrid-linear architecture (63 KDA and 9 Gated MLA layers, 512 routed experts, 8 active per token) aimed at agents, search, office productivity and software engineering. A later post-trained snapshot, Ling-3.1-flash-1019, lifts the Artificial Analysis Intelligence Index v4.3.2 score from 41 to 44 according to its model card. The card lists SGLang (Docker image lmsysorg/sglang:dev-ling-3.1-flash, 8x H200) and vLLM as serving options.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 560B total / 25B active parameters, 1M-token context window
- Hybrid-linear MoE: 63 KDA + 9 Gated MLA layers, 512 routed experts, 8 activated per token
- MIT license with weights on Hugging Face
- Model-card results include SWE-bench Pro 68.74 and Terminal-Bench 4.0 46.03 (starred, unexplained on the card)
- Served with SGLang or vLLM using the ling3 tool-call and reasoning parsers
Use cases
- Self-hosted agentic coding and terminal workflows
- Long-context search and office-productivity agents
- Research on hybrid linear-attention MoE architectures
Pricing
Open weights under MIT. Hosted by third parties: OpenRouter lists a free tier with a 262K context window; DeepInfra and Vercel AI Gateway list paid per-token pricing (rates not verified).
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- Not listed
- Record updated
- 2 Nov 2026
- Record last checked
- 2 Nov 2026