inclusionAI

Every Ling 3.0 model inclusionAI has shipped in 2026 reached a commercial API before it reached Hugging Face, and the newest one has not reached Hugging Face at all. The lab is Ant Group's open-source arm — "inclusionAI (IAI), founded by Ant Group with researchers from OpenAI, Google, and FAIR (Meta AI)" — the About Team section; the page footer reads "© 2026 inclusionAI, Ant Group. All rights reserved."source, accessed 2026-09-06 — and it states the commitment plainly on its own site, as one of three mission cards: "Share code, models, and data (within legal/ethical guidelines) to accelerate global AGI progress" — the Mission card headed "Open-Source Commitment"source, accessed 2026-09-06. It keeps that commitment. What the record shows is that it keeps it second.

The launch of Ling-3.0-flash is the clearest instance because the interval is documented on both ends. On 23 July the account posted ten tweets posted between 17:55:09Z and 17:56:55Z on 2026-07-23 — three on specifications and availability, seven captioned "Demo 1" through "Demo 7" — none of which mentions weights, a licence, or Hugging Facesource, accessed 2026-09-06. The specification tweets are confident and precise — "124B parameters. Just 5.1B active per token. With 1/8 of the total and 1/12 of the active parameters, it matches or beats our 1T flagship model on most benchmarks shown." — the opening tweet of the launch thread, 2026-07-23T17:55:09Zsource, accessed 2026-09-06, on "Ling-3.0 starts with native hybrid-linear attention: KDA and MLA layers stacked 5:1." and "It supports 256K context natively and can scale to 1M." — the second tweet of the launch threadsource, accessed 2026-09-06 — and the third tweet put the model on OpenRouter free through 3 August. Nineteen minutes after the first tweet, a reply on the thread asked "any open weights? I noticed you forgot the hugging face link..." — a reply on the launch thread timestamped 2026-07-23T18:14:41Z, nineteen minutes after the opening tweetsource, accessed 2026-09-06. The account answered the next morning, in a post about inference partners: "Free access runs through Aug 3. Open-source release coming soon—stay tuned." — a follow-up tweet timestamped 2026-07-24T07:22:34Z, the first mention of open weights by the account itselfsource, accessed 2026-09-06. The repository appeared on 2 August, ten days after the thread, and it was mit — the `cardData.license` and the `license:mit` tag on inclusionAI/Ling-3.0-flashsource, accessed 2026-09-06 — not the Apache 2.0 that llm-releases had recorded from the same announcement, and that its page still reports today alongside "Open weights · Apache 2.0 (announced; weights not yet posted as of 2026-07-24)" against a Weights row reading "Not released", still served on 2026-09-06source, accessed 2026-09-06. The weights have been downloadable for five weeks; the tracker has not noticed.

The pattern held twice more and the gap is closing. Ling 3.0 Flash Fin went up on the router on 27 August and its weights followed on 3 September — seven days, down from ten. Ling 3.0 Flash Sante went up on 4 September, and as of today no repository whose name contains "Sante" exists under the inclusionAI account — a Hugging Face model search for that string, scoped to the account, returns an empty listsource, accessed 2026-09-06; a direct request for inclusionAI/Ling-3.0-flash-Sante returns an authentication error rather than a record. The counter-example is the one model that was never sold: Ling-3.0-tiny went straight to Hugging Face on 10 August under MIT and has no router row at all. Weights-first happens here when there is no API to lead with.

Every inclusionAI row in the OpenRouter snapshot of 5 September 2026 is one model wearing three names. Fin declares its parent outright — inclusionAI/Ling-3.0-flash, declared as `cardData.base_model` and as the tag `base_model:finetune:inclusionAI/Ling-3.0-flash` on inclusionAI/Ling-3.0-flash-Finsource, accessed 2026-09-06 — and the weights agree to the digit: 127,486,405,600 (F32 165,472 plus BF16 127,486,240,128) — the same safetensors totals the API reports for the base model, digit for digit, across 64 shards rather than 24 plus a separate model-mtp shardsource, accessed 2026-09-06, against 127,486,405,600, the safetensors parameter total the Hugging Face API reports for inclusionAI/Ling-3.0-flash (F32 165,472 plus BF16 127,486,240,128)source, accessed 2026-09-06 on the base. A fine-tune and a set of finance tags are the whole difference between them. The catalog does not list them alike: in that snapshot inclusionai/ling-3.0-flash heads at $0.02per million tokensopenrouter-models, last checked 2026-09-14 and inclusionai/ling-3.0-flash-fin heads at $0.06per million tokensopenrouter-models, last checked 2026-09-14. Each of those is the top listed provider's rate for its own row rather than necessarily inclusionAI's own, and nothing requires the two rows to be headed by the same provider.

What is not a listing artefact is how much room the two Fin tiers give an answer. They are the same weights behind the same declared context length, and `top_provider.max_completion_tokens` of 235,929 on the paid Fin row against 32,768 on its free twin, from identical `context_length` of 262,144source, accessed 2026-09-06 — identical context in, a seventh of the room to reply.

Worth knowing before trusting the headline number: the shipped file is larger than the announcement. The launch said 124B parameters; the safetensors index reports 127,486,405,600, the safetensors parameter total the Hugging Face API reports for inclusionAI/Ling-3.0-flash (F32 165,472 plus BF16 127,486,240,128)source, accessed 2026-09-06.

Distribution went somewhere else entirely. Two days after the weights landed, AtomicChat/Ling-3.0-flash-GGUF, created 2026-08-04T20:00:45Z — two days after the weights — with 265,904 downloads, 59 likes, and thirty quantisation build directories from BF16 down to AD-IQ1_S plus an imatrixsource, accessed 2026-09-06 — a stranger's repository, and a more thorough one than the lab ever shipped. inclusionAI got to its own GGUF on 31 August, twenty-seven days later: inclusionAI/Ling-3.0-flash-GGUF, created 2026-08-31T03:18:15Z, with 427 downloads, 4 likes, and five build directories — BF16, Q4_K_M, Q5_K_M, Q6_K, Q8_0source, accessed 2026-09-06. Against the same base weights at 396 likes and 18,655 downloads on inclusionAI/Ling-3.0-flash, repository created 2026-08-02T16:14:41Zsource, accessed 2026-09-06, the community conversion has been fetched more than six hundred times as often as the vendor's. And Ling is not even where this org's downloads are: its most-fetched artefact is inclusionAI/LLaDA2.0-mini at 229,915 downloads against 70 likes, ahead of LLaDA2.1-mini at 111,105 — both diffusion language models; the most-liked repository is Ling-1T at 545 likes and 2,679 downloadssource, accessed 2026-09-06. The whole of this lab's presence in the 5 September 2026 catalog snapshot is one fine-tuned MoE, while its actual traffic sits in a diffusion line no router carries at all.