Ling-3.0-flash-base: o lançamento discreto de checkpoint do Ant Group, agora oficial

Ant Group's inclusionAI lab pushed Ling-3.0-flash-base and five sibling checkpoints to Hugging Face on 11 August 2026 with no announcement attached, and for a full week all six repositories sat at zero downloads — nobody outside the lab appeared to have noticed. Then, on 19 August, the lab's official account publicly introduced the collection, and every model card in it was updated within minutes. This is a what-we-know-so-far piece about that release. The short version up front: Ling-3.0-flash-base is not a chat model, and it is not meant to be one. It is the raw, pre-alignment base checkpoint of Ling-3.0-flash, the 124B-parameter Mixture-of-Experts model with 5.1B active parameters, released under MIT for people who want to keep training it rather than call it.

The timeline: a week of silence, then an announcement

The sequence matters, because it tells you how to read everything else in this piece. The six repositories were all created on 11 August, and the weights for Ling-3.0-flash-base — the merged checkpoint, the one this article is named for — landed with an "initial model release" commit on 12 August. No blog post, no social thread, no announcement accompanied any of it. At research time, a week later, every one of the six repos still showed zero downloads.

That silence ended on 19 August. The official Ant Ling account announced the checkpoint collection publicly, and the Hugging Face cards for all six repos were updated within a few minutes of each other — the timestamps line up almost exactly. So the honest framing is: the weights shipped quietly, the announcement came a week late, and everything currently knowable about the release comes from the model cards themselves plus a short training-cookbook repository. Nothing else has been independently verified yet.

What actually shipped: six checkpoints, three training stages

InclusionAI describes the release as "a collection of checkpoints during the training process" for the Ling-3.0 series, its most efficient language foundation model family to date. For each of the two smallest sizes — Ling-3.0-tiny and Ling-3.0-flash — the lab published three checkpoints, one per training stage:

• Pre-trained — Ling-3.0-flash-base-30T and Ling-3.0-tiny-base-30T: completed large-scale pre-training; no mid-training, no merging, no post-training.

• Mid-trained — Ling-3.0-flash-base-midtrain and Ling-3.0-tiny-base-midtrain: completed the mid-training phase; no checkpoint merging and no post-training.

• Merged — Ling-3.0-flash-base and Ling-3.0-tiny-base: the WSM-merged base checkpoints, produced from the mid-trained weights and representing the final base state before post-training.

All six are base checkpoints. None has gone through the alignment stage — supervised fine-tuning, preference optimization, RL — that turns a base model into something you can put in front of users. The model card says so in plain terms: not recommended for direct end-user chat deployment, safety-critical applications without additional alignment and evaluation, or production use without post-training and task-specific validation.

What a base checkpoint is — and why the WSM twist matters

If you have only ever consumed models through an API, a "base checkpoint" can sound like a smaller or older version of a model you already know. It is neither. A base checkpoint is the unaligned neural-network weight file that comes out of pre-training; the instruction-following, tool-calling behavior you actually use is layered on top of it afterwards. Releasing base checkpoints is how a lab hands the research community the raw material to do its own continued pre-training, fine-tuning, distillation, and RL.

What makes this release more interesting than a typical base-weight dump is the training recipe behind it. InclusionAI is using a technique it calls Warmup-Stable and Merge (WSM), and it replaces the conventional learning-rate decay at the end of training with weighted checkpoint merging. In plain terms: instead of letting the learning rate wind down to finish the run, the lab trains at a stable rate, saves checkpoints along the final stretch, and merges them with learned weights. The vendor's stated motivations, from the model card:

• Because there is no decay phase to "bake in" the final data mix, the resulting base model is better suited to continued pre-training and dynamic data expansion.

• Because the merge recipe can be explored offline, a team can compare different decay profiles without rerunning the expensive training run for each one.

That second point is the one that makes the release a research artifact rather than a curiosity: the whole point of these checkpoints is that they are meant to be trained on further, and the recipe is designed to tolerate it.

What's inside Ling-3.0-flash-base

The architecture is identical to the post-trained Ling-3.0-flash, which should already be familiar from the wider coverage of that model. From the model card and config.json:

• Size — 124B total parameters per the vendor (the full checkpoint on disk, including embeddings, counts to roughly 127B on Hugging Face), with 5.1B activated per token.

• Sparsity — a 1/64 sparse MoE: 512 routed experts, with 8 routed experts and 1 shared expert activated for every token.

• Attention — a native hybrid-linear design: 35 Kimi Delta Attention (KDA) layers and 7 Gated MLA layers in a 5:1 repeating pattern across 42 transformer layers.

• Context — the config carries a 256K-token position embedding (262,144), the same native context as the post-trained sibling.

• Weights — BF16, shipped as 28 safetensors shards (~255 GB on disk), with the custom bailing_hybrid modeling code bundled in the repository.

• License — MIT, with no acceptable-use rider.

One honest caveat about the size figure: the vendor's model card says "Total 124B," and that is the number that circulates everywhere — but the total safetensors parameter count on the repository is about 127.5B once the embedding layer is included. The 124B figure is the non-embedding total, and the two numbers are not in conflict; it is worth knowing which one you are looking at when you see "127B" in a sidebar or a downloader.

What it is for, and what it is not

The model card's recommended use cases are unambiguous:

• Continued pre-training

• Mid-training

• Supervised fine-tuning for domain adaptation

• Preference optimization and RL post-training

• Distillation research

• Long-context and MoE systems research

And its "not recommended as-is" list is just as explicit: direct end-user chat deployment, safety-critical applications without alignment, production use without post-training. In other words, this is raw material for people who train models, not a model people deploy. If you have no plans to fine-tune or pre-train anything, there is nothing for you to download here — and nothing wrong with that.

The other half of the story is the scale-seamlessly design. InclusionAI states that Ling-3.0-tiny-base and Ling-3.0-flash-base share the same training recipe, so the community can validate a continued-pretraining or fine-tuning strategy on the tiny checkpoint first — which fits in far less GPU memory — and then scale the same validated recipe to the larger flash checkpoint. That is a deliberate workflow choice, not an accident of release logistics, and it is the most useful property of the collection for anyone who actually wants to experiment.

Hardware reality, and the route most people actually want

There is no getting around the size of the flash base weights. BF16 at ~255 GB means a serious multi-GPU node (or a memory-optimised training setup) before you even start; there is no official FP8 or FP4 quantisation of the base checkpoint — the quantised variants that exist in the org's catalogue are for the post-trained model. The fine-tuning examples live in the inclusionAI/ling-cookbook GitHub repository, which is explicitly a work in progress. If continued pre-training is not your team's focus, the base checkpoints are not your entry point.

Most people reading about the Ling-3.0 family do not want a base checkpoint at all — they want the model that talks. That is the post-trained Ling-3.0-flash, available through the vendor's own API and several third-party platforms. When a fast, newly released open-weights model like that appears, the cheapest way to judge it against your current stack is to run it side by side with what you already use — and that is exactly what a router is for: one API, one key, and automatic failover so a provider hiccup does not end the trial. OrcaRouter connects to 200+ models this way and passes provider list prices straight through at 0% markup, so the price you see is the provider's own price, and a vendor cut lands on our side the same day.

What's not yet confirmed

Because this release is a day old in announcement terms, the unconfirmed list is long and should be read as a to-do list, not a criticism:

• No independent benchmarks of the base checkpoint. The model card embeds a vendor benchmark chart comparing Ling-3.0-flash-base with pretrained peers across math, coding, reasoning, multilingual, and long-context suites — but it is the vendor's own evaluation, hosted on an internal image server, with no methodology section and no third-party reproduction.

• No community reproductions yet. As of research time, all six repos showed zero downloads and essentially no engagement. The release is too fresh for anyone to have published an independent fine-tune or continued-pretraining run on top of these weights.

• Tooling is early. The bailing_hybrid architecture ships with custom modeling code, so you are depending on trust_remote_code or on a framework that has added support. The fine-tuning cookbook is explicitly work in progress.

• The announcement itself is thin. The public introduction came as an X thread and the model cards; there is no standalone launch document describing the training data, the compute, or the exact merge recipe beyond what the cards state.

By contrast, the post-trained sibling does have an independent track record: Ling-3.0-flash is tracked by Artificial Analysis at an Intelligence Index of 38 — a figure that comes from the leaderboard, not from the vendor. None of that exists yet for Ling-3.0-flash-base; a base checkpoint is not served anywhere, so it is not on any leaderboard. The two artifacts are at completely different stages of verification, and it helps not to blur them.

Quem deve agir agora

Three groups should read this differently:

• Researchers doing continued pre-training, domain SFT, or distillation — this is the release for you. The WSM-merged base checkpoint is specifically built to be trained on further, and the tiny-then-flash recipe gives you a cheap way to validate an approach before scaling it. Start with Ling-3.0-tiny-base, prove the recipe, then move to Ling-3.0-flash-base.

• Engineers who want a working model — you want the post-trained Ling-3.0-flash, not this release. The base checkpoints will not follow instructions and are not served anywhere; do not point an API client at them.

• Everyone else — the useful move is to watch. The interesting signals over the next few weeks are a first independent fine-tune built on these weights, a community reproduction of the vendor benchmark chart, and whether a third-party training stack adds first-class support for the bailing_hybrid architecture.

Perguntas Frequentes

What is the difference between Ling-3.0-flash-base and Ling-3.0-flash?

Ling-3.0-flash is the post-trained model — aligned with supervised fine-tuning and RL, follows instructions, does tool-calling, and is served through APIs. Ling-3.0-flash-base is the same architecture at an earlier stage: it is the final merged base checkpoint before any post-training, and it will not behave like a chat model. The base exists so that researchers can do their own continued pre-training, fine-tuning, or distillation rather than starting from scratch.

Why would a lab release checkpoints that have not been post-trained?

Because post-training is exactly what most researchers want to do themselves, and a base checkpoint is the correct starting point for it. Releasing the pre-trained, mid-trained, and merged checkpoints separately means a team can pick the stage that matches their work: continued pre-training from the pre-trained weight, or fine-tuning and RL from the merged base — and, thanks to WSM, without the learning-rate-decay constraints that make a base model fragile when you train it further.

Can I run Ling-3.0-flash-base on consumer hardware?

Not practically. The BF16 weights are roughly 255 GB across 28 shards, with no official quantised variant of the base checkpoint. The smaller sibling, Ling-3.0-tiny-base, is the one intended for experimentation on more modest hardware — and it shares the flash checkpoint's training recipe, so validated experiments scale up.

Is Ling-3.0-flash-base free to use commercially?

Yes — it is released under the MIT license with no acceptable-use rider. That covers both research use and, in principle, commercial use of the weights and of any fine-tunes built from them. It is a base checkpoint, so "use" here means training on it; the license says nothing about a hosted API, because none exists for the base model.

Conclusão

For most readers, Ling-3.0-flash-base is a pointer, not a product: it tells you where the Ling-3.0 family is headed and how the lab intends the community to build on it, but there is nothing here to call through an API. For the smaller group actually doing continued pre-training, fine-tuning, or distillation, it is a genuinely useful release — six checkpoints, MIT-licensed, deliberately designed to be trained on further, with a tiny-first path that makes experimentation affordable. What would move it from "useful release" to "proven recipe" is exactly what is missing today: an independent reproduction, by someone outside the lab, of what WSM checkpoints can do when somebody keeps training them.