inclusionAI/Ling-3.1-flash is a public Hugging Face repository of weights. The Hub API records createdAt 2026-10-13T11:16:41.000Z and lastModified 2026-10-15T02:30:27.000Z, with commit a4ccd6b45c002ddb236f9cb77d52cc427482bf19. The organization page calls inclusionAI the home of Ant Group’s artificial general intelligence work, and it points at Robbyant as another organization under Ant Group. Wikipedia’s Ant Group page carries the short description “Chinese financial services company” and the category “Companies based in Hangzhou.”
The model card’s YAML sets license: mit. The same value is cardData.license on the model API. The repository’s file listing had no LICENSE file on 15 October 2026. The MIT declaration read here is metadata on the card, not a license text file in the repo. The organization page says its projects use MIT or Apache licenses. That is a general line, not the missing file.
What the card states
The card introduces Ling-3.1-flash as an upgraded native hybrid reasoning model built on Ling-3.0-flash. It states 560 billion total parameters, 25 billion active parameters per token, and a 1 million token context window. Its overview table lists a hybrid-linear mixture of experts: 63 KDA layers plus 9 gated MLA layers in a 7:1 ratio, 4 dense layers, 512 routed experts, 1 shared expert, 8 activated experts, 64 attention heads, hidden size 4096, expert intermediate size 1280, dense intermediate size 10240, and vocabulary size 157,184.
The card names GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, Terminal-Bench 4.0, SWE-bench Pro, Automationbench, MCP-Atlas and SkillsBench, and it says the model is used inside Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw. The numeric charts are images. This story does not quote scores from those images. The card also describes LingCode, a harness on the coding agent Pi, and says inclusionAI is considering open-sourcing it later. LingCode is not in this repository.
What config.json records
config.json names architecture BailingMoeV3ForCausalLM and model_type bailing_hybrid. It records 72 hidden layers, first_k_dense_replace 4, 512 experts, 8 experts per token, 1 shared expert, 64 attention heads, hidden size 4096, expert intermediate size 1280, dense intermediate size 10240 and vocabulary size 157,184. Those counts match the card’s table. The card’s 63 KDA layers plus 9 gated MLA layers add to 72, which is the config’s hidden-layer count. The file does not itself label the 63/9 split.
max_position_embeddings is 262,144, and rope_scaling is null. 262,144 is 256 times 1,024. The card’s 1 million token window is not this field. torch_dtype is bfloat16. A quantization_config block sets method fp8, format e4m3, with a list of modules left unconverted, including the language-model head and word embeddings.
The Hub API’s safetensors metadata sums stored parameters to 572,120,680,768, of which 570,628,505,600 are recorded as FP8 E4M3, 1,491,619,712 as BF16 and 555,456 as F32. That total is the stored-tensor sum on the API. It is not the card’s 560 billion figure, and it is not an active-parameter count. The 25 billion active figure appears on the card, not in config.json.
Recipes on the card
The card’s recommended path is SGLang, with image lmsysorg/sglang:dev-ling-3.1-flash. Its server example uses 8 H200 GPUs with tensor parallel 8 and expert parallel 8, and it says 4 GB300 GPUs can use tensor parallel 4 and expert parallel 4. A vLLM section clones the vllm branch of vLLM and serves with tensor parallel 8, expert parallel, tool parser ling3 and reasoning parser ling3. Those are inclusionAI’s recipes. This story does not report a run of either one. The card points to an SGLang cookbook for further measured accuracy and speed numbers. That cookbook was not opened here, so those measurements are not repeated.
The trial, as Pandaily tells it
The model card read here does not describe an API trial. Pandaily, with a CMS publishedDate of 2026-10-15T03:50:00.000Z, says the model was announced on 30 September with a two-week free API trial capped at a 256K-token context, that the team then said it planned to open the full 1 million token context and open-source the weights after the free period, and that the free period closed on 14 October as the weights went up. The 30 September announcement was not opened for this story. Pandaily’s 256K cap is the same token count as max_position_embeddings in the config file opened on 15 October. The Hub API’s createdAt for the repository is 13 October, a day before the date Pandaily gives for the weights going up.
Update, 27 October 2026: Ling-3.1-flash-1019 and a base model
inclusionAI has posted Ling-3.1-flash-1019, “an upgraded version” of Ling-3.1-flash with the same 560B total and 25B active parameters. The card again declares MIT, and again the repository has no separate LICENSE file. The model is not gated.
The card says further post-training improved reasoning and agent tasks. Its figures, attributed to Artificial Analysis:
| Benchmark | Ling-3.1-flash | 1019 |
|---|---|---|
| AA Intelligence Index v4.3.2 | 41 | 44 |
| AA-Briefcase (Elo) | about 1400 | 1524 |
| AutomationBench-AA | 62% | 68% |
| CritPt | 18% | 26% |
| SciCode | 54% | 57% |
The same context mismatch remains: the card says 1M tokens, while config.json in the 1019 repository still sets max_position_embeddings to 262,144.
inclusionAI also posted Ling-3.1-flash-base, the pretrained checkpoint, under MIT on 25 October, which is useful if you want to run your own fine-tune.

The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment