by inclusionAI
Open-weights 560B hybrid-linear MoE reasoning model (25B active, 1M context) from inclusionAI, MIT licensed.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Ling-3.1-flash is a native hybrid reasoning language model from inclusionAI, an upgrade of Ling-3.0-flash. It uses a hybrid-linear mixture-of-experts architecture with 560B total parameters and 25B active per token, 512 routed experts (8 active) plus one shared expert, and a layer stack of 63 KDA linear-attention layers and 9 gated MLA layers. The context window is 1M tokens.
The model card targets general-purpose agents, search, office productivity and software engineering, and reports results on benchmarks such as Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA V2.1, HealthBench Professional, MCP-Atlas and SkillsBench. The vendor says it works with agent frameworks including Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw. It also describes LingCode, a harness built on the Pi coding agent for long-horizon coding tasks, which is not released. A YouTube reviewer described its performance as close to leading open models, but these claims are not independently verified here.
Weights are published on Hugging Face under the MIT license in FP8 (roughly 574 GB, per one reviewer). Self-hosting is documented with SGLang (day-0 image, MTP/NEXTN speculative decoding; 8×H200 or 4×GB300) and vLLM. The model is also listed on OpenRouter and ModelScope. Running it locally requires multi-GPU datacenter hardware.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open model with 560B total and 25B active parameters, performing close to leading open models like GLM 5.3, Kimi and DeepSeek 4.1. About 574 GB.
Closest neighbours by meaning, not just by category.
Model · inclusionAI
Open-weights 124B MoE healthcare model (5.1B active) from inclusionAI, specialized from Ling-3.0-flash.
Model · Ant Group
Ant Group's multimodal mixture-of-experts model (560B total, 25B active) in the Ling family, served via API.
Model · Baidu
Baidu's open-weights 280B-parameter sparse MoE base model (14B active) with 1M-token context, released pre-post-training.
Model · Rednote
Open MoE language model from Rednote (Xiaohongshu) with 208B total and 16B active parameters.
Model · Alibaba
Alibaba's Qwen3.8 Flash Next: a 125B-parameter MoE language model with about 6B active parameters, previewing Qwen 4 design.
Model · MiniMax
Open-weights native multimodal MoE model from MiniMax, ~428B parameters (~23B active) with 1M-token context.