MITllmRuns locally560B
Ling-3.1-flash: Specs, License & How to Run
inclusionAI MIT 560B/25B hybrid-linear MoE with official SGLang and vLLM recipes.
- License
- MIT
- Parameters
- 560B (active 25B)
- HF likes
- 121 (2026-10-19)
- First seen
- 2026-10-13
TL;DR. Ling-3.1-flash is a llm model from inclusionAI (Ant Group). Parameters: 560B (active 25B).License: MIT. It is listed under ChatGPT, Closed LLM APIs, Claude Code.
| Developer | inclusionAI (Ant Group) |
|---|---|
| Type | llm |
| Architecture | Hybrid-linear MoE (BailingMoeV3ForCausalLM); 72 layers (README 63 KDA + 9 Gated MLA); 512 routed experts, 8 active, 1 shared; hidden size 4096; vocab 157184 |
| Parameters | 560B · active 25B |
| Modalities | text |
| Context | 1048576 (README); 262144 (config.json max_position_embeddings) |
| License | MIT — Hugging Face cardData.license is mit and the Hub tags include license:mit as of 2026-10-19. This check found no LICENSE file in the repository (raw LICENSE URL 404). Bundled modeling_bailing_moe_v3.py carries an Apache-2.0 header. Not legal advice. |
| Hardware | Not yet verified — Card README (opened 2026-10-19) gives a day-0 SGLang recipe on 8× H200 with TP8 + EP8, or 4× GB300 with TP4 + EP4. Hugging Face usedStorage is 573,813,820,876 bytes (~534.5 GiB on disk); safetensors.parameters total 572,120,680,768. Peak VRAM is not independently measured. Datacenter-class. (not independently tested) |
| Hugging Face likes | 121 (as of 2026-10-19) |
| Model card | https://huggingface.co/inclusionAI/Ling-3.1-flash |
| First seen | 2026-10-13 |
Ling-3.1-flash is an open-weight text MoE from inclusionAI. Official hosts we opened on 2026-10-19: inclusionAI/Ling-3.1-flash, the README, config.json, inclusionAI/ling-cookbook (33 stars), the SGLang cookbook, and the site ant-ling.com.
This page is a spoke from ChatGPT alternatives, closed LLM APIs, and Claude Code alternatives. Those hubs name product classes. We do not claim ChatGPT or Claude Code quality.
What we verified (2026-10-19)
| Field | Value | Source |
|---|---|---|
| Host | inclusionAI/Ling-3.1-flash |
HF API |
| Created | 2026-10-13T11:16:41Z | HF createdAt |
| Last modified | 2026-10-15T02:30:27Z | HF API |
| Gated | false | HF API |
| Likes / downloads | 121 / 2,191 | HF API |
| License | mit | HF cardData / tags |
| LICENSE file | 404 | .../resolve/main/LICENSE |
| Pipeline tag | text-generation | HF API |
| Architecture class | BailingMoeV3ForCausalLM |
HF config |
| safetensors total | 572,120,680,768 | HF API |
| usedStorage | 573,813,820,876 bytes | HF API |
| Cookbook stars | 33 | GitHub API inclusionAI/ling-cookbook |
README overview (copied): 560B total / 25B activated, 72 layers (63 KDA + 9 Gated MLA), 4 dense layers, 512 routed experts, 1 shared, 8 activated, hidden size 4096, vocab 157184, 1M-token context. config.json in the same repo sets max_position_embeddings to 262144, num_hidden_layers 72, num_experts 512, num_experts_per_tok 8, vocab_size 157184. SGLang’s cookbook also quotes a 262,144-token limit. We keep both numbers.
How to run (from official README, not yet tested)
Day-0 SGLang image named on the card: docker pull lmsysorg/sglang:dev-ling-3.1-flash. Server example is 8 × H200, --tp 8 --ep 8, or 4 × GB300, --tp 4 --ep 4. Sampling copied from the card: temperature=1.0, top_p=0.95, top_k=20, enable_thinking true. vLLM is a second official path (tool-call-parser ling3, reasoning-parser ling3). forkalt has not served the checkpoint. See also DeepInfer vs vLLM if you are choosing a DeepSeek-family engine rather than this Ling recipe.
License
Card YAML license: mit. No LICENSE file in the repo on this check. Bundled modeling code states Apache-2.0. Not legal advice.
What we did not invent
Vendor benches (GDPVal-AA V2.1, HealthBench Professional, FrontierSWE, Terminal-Bench 4.0) stay on the card. The README’s “strong user experience across Claude Code, Kilo Code, …” sentence is upstream. We have no sourced bake-off. Hardware VRAM is not yet verified.
Related chat weights: DeepSeek-V4.1-Flash, Reflection Beam, Kimi K3. Runtimes: vLLM, DeepInfer. Coding agents: Cline, OpenCursor.
Deploy options
- local: Hugging Face weights
- local: SGLang day-0 image
- local: vLLM recipe
- local: GitHub cookbook
FAQ
Are Ling-3.1-flash weights public?
Yes. huggingface.co/inclusionAI/Ling-3.1-flash is ungated (created 2026-10-13T11:16:41Z; last modified 2026-10-15T02:30:27Z; 121 likes / 2,191 downloads in the 2026-10-19 API check; pipeline_tag text-generation; card license mit). The Hub file list includes 54 safetensors shards and no LICENSE file.
How big is it?
The README model overview (checked 2026-10-19) lists 560B total / 25B activated, 512 routed experts with 8 active plus 1 shared, and a 1M-token context window. Hugging Face safetensors.parameters total 572,120,680,768 (~572.1B). config.json sets max_position_embeddings to 262,144. We copy both figures; we do not pick a winner.
Can I use it commercially?
The card declares MIT. There was no LICENSE text file in the repo on 2026-10-19. Modeling code in the same repo carries an Apache-2.0 header. Not legal advice; read the card and the Python headers.
Does it match ChatGPT or Claude Code quality?
No. This page identifies an open-weight checkpoint. We have no sourced bake-off.
Sources & verification
- https://huggingface.co/inclusionAI/Ling-3.1-flash
- https://huggingface.co/inclusionAI/Ling-3.1-flash/raw/main/README.md
- https://huggingface.co/inclusionAI/Ling-3.1-flash/raw/main/config.json
- https://github.com/inclusionAI/ling-cookbook
- https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.1-flash
- https://recipes.vllm.ai/inclusionAI/Ling-3.1-flash
- https://ant-ling.com
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.