inclusionAI releases Ling 3.1 Flash open weights
A 560B-parameter MIT-licensed MoE with a 1M-token context
Confirmed
The takeaway
Two weeks after its API-only free trial, Ant Group’s inclusionAI published the weights of Ling 3.1 Flash on Hugging Face under the MIT license (repo created Oct 13, files finished Oct 14, 2026).
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 14 October 2026
Your AI and this story
- GPT-6 Astra167 days after its cutoff
- Claude Opus 5.5106 days after its cutoff
- Gemini 4 Argon197 days after its cutoff
- Grok 4.7136 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 106 days before it.
Key facts
- Hugging Face: inclusionAI/Ling-3.1-flash, license MIT, 54 safetensors shards; repo created 2026-10-13, last modified 2026-10-14 (HF API)
- Architecture (model card): hybrid-linear MoE, 560B total / 25B active, 72 layers = 63 KDA (linear attention) + 9 gated MLA layers (7:1), 512 routed experts + 1 shared, 8 active, vocabulary 157,184
- Context window: 1M tokens (the trial was capped at 256K)
- Model card names GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, Terminal-Bench 4.0, SWE-bench Pro, MCP-Atlas and SkillsBench as strong results; scores are only in images, not text
- Serving: day-0 SGLang image (8x H200 TP8+EP8, or 4x GB300) with MTP speculative decoding; vLLM branch also supported
- Also introduces LingCode, a long-horizon coding harness built on the open-source Pi agent; inclusionAI is ‘considering open-sourcing it in the future’
- A training-content summary (TDS) PDF is published in the inclusionAI/AI-Transparency repo
What happened
Ling 3.1 Flash launched around Sept 29–30, 2026 as a two-week free trial on API gateways, with a promise to open-source it afterwards.
The trial ended around Oct 13. On Oct 13–14 inclusionAI published the full weights on Hugging Face (inclusionAI/Ling-3.1-flash,
54 safetensors shards, MIT license), with a model card giving the architecture: a hybrid of 63 Kimi-style delta attention (KDA) linear
layers and 9 gated MLA layers, 512 routed experts with 8 active, 25B active parameters out of 560B, and a 1M-token context.
The card claims strong results on agentic and office benchmarks (GDPVal-AA V2.1, Terminal-Bench 4.0, FrontierSWE, SWE-bench Pro and others) but gives the numbers only as images, so they are not quoted here. It also announces LingCode, a coding-agent harness built on the open-source Pi agent, which is not yet released. No independent evaluations had appeared by the evening of Oct 14.
Why it matters
inclusionAI kept its promise on time: the API-first, weights-later pattern was followed through, unlike StepFun’s Step 5 Preview, whose Oct 15 weights were still not public on Oct 14. A 560B model under the permissive MIT license, with a 1M-token context and a mostly linear-attention design, gives open-model users a large long-context option. Its claimed scores are unverified until outside tests appear.
Sources
5 sources from 3 sites. Numbers match the chips in the text.
5 sources: 4 primary, 1 press
Primary
- Hugging Face: inclusionAI/Ling-3.1-flash (model card)huggingface.co, code
- Hugging Face API: repo metadata (created/modified dates, license, files)huggingface.co, docs
- SGLang cookbook: Ling-3.1-flashdocs.sglang.io, docs
- inclusionAI: Ling-3.1 training-content summary (PDF)huggingface.co, docs
Press
Changes
- Filed (HF repo and model card read; outcome of upcoming item inclusionAI releases Ling 3.1 Flash open weights)