Ant Group's inclusionAI open-sources Ling-3.1-flash under MIT: 560B hybrid-linear MoE (25B active), 1M context, aimed at agents and coding

Why it matters

【Key points】

Ant Group's inclusionAI has open-sourced Ling-3.1-flash on Hugging Face (repo created Oct 13, still being updated Oct 15) under the MIT license, ungated. Per the model card: a native hybrid reasoning model with 560B total and 25B active parameters per token and a 1M-token context. The architecture is a hybrid-linear MoE (63 KDA + 9 Gated MLA layers, 512 routed experts with 8 active), building on Ling-3.0-flash. It targets general agents, search, office work and software engineering, claims strong results on Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE and GDPVal-AA V2.1, and says it works in Claude Code, Qwen Code, Hermes Agent and similar frameworks. Weights ship in FP8 (~572B params in safetensors). SGLang has a day-0 image (8×H200 or 4×GB300), vLLM is supported, and ModelScope and OpenRouter are live. The team also mentions LingCode, an in-house long-horizon coding harness it may open-source later.

【Verifiable】

Hugging Face: https://huggingface.co/inclusionAI/Ling-3.1-flash (createdAt 2026-10-13T11:16Z, license mit, gated=false)

OpenRouter: https://openrouter.ai/inclusionai/ling-3.1-flash

ModelScope org: https://modelscope.cn/organization/inclusionAI

X news card (discussed since Oct 14; ~115 Ling-3.1 posts on Oct 14, ~60 on Oct 15 through 06:27Z): https://x.com/i/trending/2110411771840548958

【Our read】

1. This is another 500B-class, permissively licensed open-weight model from a major Chinese company. Alongside StepFun's Step 5 Preview weights the same week (e/876), it undercuts the "Chinese labs have gone quiet" narrative.

2. The model card gives its benchmark results only as images, with no number table. Claims on X like "210 tok/s" and an "805-line app" come from community demos, and no independent third-party evaluation exists yet.

3. Running it takes about 8 H200-class GPUs, so "easy local use" is overstated. For most people the realistic route is the OpenRouter API.