News · 2026-08-05

Ant published Ling-3.0-flash's weights under plain MIT, with no rider attached

Ant Group's InclusionAI lab has published the complete weights for Ling-3.0-flash, a 124-billion-parameter mixture-of-experts model, on Hugging Face and ModelScope under an unmodified MIT license. There is no acceptable-use policy, no revenue threshold, no branding clause, and no field-of-use carve-out anywhere in the release. When Ground Truth covered this model on July 23, it was API-only with no published checkpoint and no clear license, and we said it should be treated that way until Ant published weights. Ant has now published them.

Key facts

• 124 billion total parameters with about 5.1 billion active per token, roughly a 4 percent activation ratio.
• Two variants: Ling-3.0-flash in BF16 at about 230 gigabytes across 24 shards, and Ling-3.0-flash-fp8 at about 128 gigabytes.
• Repository timestamps put the BF16 upload at August 6, 2026 and the FP8 variant at August 5.
• License is the bare tag license: mit, matching the plain MIT text InclusionAI uses in its Ling-V2 GitHub repository.

Licensing is where most open model releases quietly stop being open, so it is worth being precise about what was checked. Both Hugging Face repositories carry license: mit in the model card front matter and nothing else in that block. Neither repository contains a LICENSE file, a USAGE_POLICY document, or any acceptable-use text; the full file tree is a gitattributes file, the README, config and model code, tokenizer files, and the weight shards. The 172-line README covers architecture, benchmark methodology, and quickstart instructions for two serving engines, and the word license appears nowhere in its body. ModelScope's record for the same model reports MIT with an empty license link and empty license name, which is the same story from a second registry.

The missing LICENSE file is a documentation gap rather than a hidden condition. A rider needs a document to state its terms, and no such document exists in the weights repository, the README, the mirror, or the lab's sibling GitHub repositories. When InclusionAI does write the grant out in full, in its actively maintained Ling-V2 repository, it is the unmodified MIT template: use, copy, modify, merge, publish, distribute, sublicense, and sell, subject only to keeping the copyright notice and the no-warranty disclaimer. One adjacent repository, their vLLM fork, is Apache-2.0, but that is inherited from upstream vLLM and applies to the serving engine, not the weights.

What is actually in the model

The architecture is more interesting than the parameter count. Ling-3.0-flash is a mixture of experts with 512 routed experts plus one shared expert, activating 8 experts per token, on top of two dense layers. The attention stack is what Ant calls native hybrid-linear: a 5-to-1 alternating pattern of Kimi Delta Attention, a linear attention variant, and gated multi-head latent attention, working out to 35 linear layers against 7 full-attention layers.

The reason to care is cost at long context. Standard attention compares every token to every other token, so doubling the input roughly quadruples the work. Linear attention approximates that comparison in a way that scales proportionally instead, at some cost in fidelity. Interleaving them five-to-one is a bet that most layers do not need the exact version, and that a periodic full-attention layer is enough to recover what the approximation loses. With a 262,000-token context window, trained on an 8K to 32K to 256K schedule, that choice is doing real work.

The release is weights plus runnable code, not just tensors: custom transformers-style modeling and configuration files ship in the same repository, and this is the post-trained chat model with a tool-calling chat template and a thinking-mode toggle, not a raw base checkpoint. It is text-only, with no vision or audio.

Why an unencumbered license is the story

Most of what gets called open weights today is not MIT or Apache. It is a bespoke community license with a revenue trigger, a naming requirement, or a use restriction that makes commercial deployment a legal question rather than an engineering one. Ant's release skips all of it, at a scale and capability level where labs usually start attaching conditions. That is a genuine competitive move as much as a philosophical one, and it lands the same week DeepSeek warned it will raise its API prices. Cheap hosted inference is exactly the thing a fully permissive checkpoint makes optional.

The honest caveat

The rollout was not clean from day one. At least one write-up published before this week's listing described the model as having no weights, no repository, and no clear license statement, which accurately described an earlier state that has since resolved. The absent LICENSE file also means the MIT grant currently rests on a metadata tag rather than a signed document in the repository, which is fine in practice and sloppy in form; a one-line file would settle it. And 230 gigabytes of weights is only meaningfully open to people who can host them, which is a smaller group than the license implies.

Primary source, verified: read the paper →

Ground Truth. (2026, August 5). Ant published Ling-3.0-flash's weights under plain MIT, with no rider attached. Ground Truth. https://groundtruth.day/news/ant-put-ling-3-0-flashs-weights-online-under-plain-mit.html

@misc{groundtruth:ant-put-ling-3-0-flashs-weights-online-under-plain-mit,
  title  = {Ant published Ling-3.0-flash's weights under plain MIT, with no rider attached},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/ant-put-ling-3-0-flashs-weights-online-under-plain-mit.html}
}

Topics: models · open-weight-models · licensing · mixture-of-experts · china · efficiency