# The gate and the giveaway

URL: https://www.thedeepfeed.ai/posts/2026-07-04-the-gate-and-the-giveaway/
Category: Models
Published: 2026-07-04
Author: the-deep-feed
Tags: open-weights, china, export-controls, coding-agents, meituan, benchmarks
Kind: deep

> While Washington spent June turning frontier model releases into licensed events, a Chinese food-delivery company open-sourced a 1.6-trillion-parameter agentic-coding model under an MIT license — trained start to finish on domestic chips, no Nvidia silicon involved. LongCat-2.0 is the counter-move to the export-control regime, and it's already downloadable worldwide.

## TL;DR

- On the July 4 weekend, Chinese food-delivery giant **Meituan** open-sourced **LongCat-2.0** — a **1.6-trillion-parameter** MoE model (~48B active) under a maximally permissive **MIT license**. It's a near-frontier **agentic-coding** model, downloadable worldwide, commercial use allowed.
- The headline fact is the hardware: Meituan says LongCat-2.0 is *"the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster"* — pretraining and inference, **no Nvidia silicon**. (It stopped short of naming the chip vendor; it credited Huawei's HCCL communication library.)
- It's not an across-the-board frontier leader — it **trails** Gemini 3.1 Pro and GPT-5.5 on reasoning benchmarks (GPQA, IFEval, BrowseComp). But it **leads them on SWE-bench Pro (59.5)**, the agentic-coding metric that actually maps to the workload buyers pay for. It spent two months topping OpenRouter usage anonymously as *"Owl Alpha."*
- The through-line to our [export-control](/posts/2026-07-09-the-voluntary-gate-that-works-like-a-license/) coverage: while Washington was turning US model releases into licensed events, China's answer was to **give the model away** — MIT, no jurisdiction, un-gateable. The gate and the giveaway are the same story from opposite ends.
- The contrarian read: the US frontier gate assumes the leverage point is *access to the best model*. LongCat says the leverage point is moving — to **good-enough agentic models that ship without a license, on chips the export regime can't touch.** A letter can't recall an MIT download.

Over the US Independence Day weekend, while Washington was busy inventing a [licensing regime for frontier model releases](/posts/2026-07-09-the-voluntary-gate-that-works-like-a-license/), a Chinese food-delivery company did the opposite. Meituan — the Beijing giant most of the world knows, if at all, as an app for ordering dinner — finished open-sourcing **LongCat-2.0**, a 1.6-trillion-parameter model, under an MIT license. Free to download. Free to run. Free to fine-tune and redistribute, commercially, anywhere, by anyone.

Two model-release philosophies collided in the same fortnight, and they could not be further apart. One turns the release into a gated event a government can throttle. The other turns it into a file on Hugging Face that no letter can recall. This is the piece about the second one, and about why the giveaway is the sharper strategic move.

# What Meituan actually shipped

Strip the geopolitics for a moment and look at the artifact. LongCat-2.0 is, in Meituan's own words on the model card, *"a large-scale MoE language model with 1.6 trillion total parameters and ~48 billion activated per token."* It has a native 1-million-token context window, a novel "LongCat Sparse Attention" mechanism, and it ships in full-precision, FP8, and INT8 variants — the INT8 checkpoint alone is 141 safetensors shards. The license, confirmed in the model-card metadata and the GitHub repo, is [MIT](https://www.opensourceforu.com/2026/06/meituan-open-sources-longcat-2-0-under-mit-license/): no regional restrictions, no usage prohibitions, no acceptable-use rider. That is more permissive than the custom community licenses DeepSeek and the Qwen family attach to their weights.

It is not a general-purpose frontier leader, and it is important to be precise about that, because the launch coverage blurred it. On the benchmark table Meituan itself published, LongCat-2.0 *trails* Gemini 3.1 Pro and GPT-5.5 on the foundational reasoning tests — GPQA-diamond, IFEval, IMO-AnswerBench, BrowseComp. Where it wins is narrower and more consequential: **agentic coding.** It posts **59.5 on SWE-bench Pro**, ahead of GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2, and it edges both on Terminal-Bench 2.1 and SWE-bench Multilingual. This is a model built to *do software work in a loop*, not to win a physics quiz.

The market had already noticed, before anyone knew what it was. For roughly two months LongCat-2.0 ran anonymously on OpenRouter under the codename **"Owl Alpha,"** quietly climbing the developer usage rankings. [VentureBeat's launch coverage](https://venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic-coding-model-thats-been-leading-openrouter-trained-entirely-on-chinese-chips) described the unveiling as "unmasking the model as the computational engine behind Owl Alpha." Developers were reaching for it on the merits, blind to its provenance, before Meituan revealed the name on June 29–30 and rolled the full weights out over the following week.

# The part that matters: no Nvidia

Here is the sentence that turns a competent coding model into a policy event. Meituan claims LongCat-2.0 is, per [SCMP's reporting](https://www.scmp.com/tech/tech-trends/article/3358854/china-debuts-biggest-ai-model-trained-local-chips-meituan-releases-longcat-20), *"the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster."*

Full-process is the load-bearing phrase. On the model card, Meituan states the whole run — *"both the full training run and the large-scale deployment"* — happened on *"AI ASIC superpods,"* spanning *"millions of accelerator-days across more than 35 trillion tokens, with no rollbacks or irrecoverable loss spikes."* The company frames it as a proof of capability: *"demonstrating that we have the capability to conduct frontier-scale training on alternative hardware platforms."*

> While DeepSeek-V4-pro relied on home-grown chips only for inference … LongCat-2.0 used domestic hardware for both inference and pre-training, according to Meituan.
>
> — [South China Morning Post](https://www.scmp.com/tech/tech-trends/article/3358854/china-debuts-biggest-ai-model-trained-local-chips-meituan-releases-longcat-20), June 30, 2026

That distinction — pretraining, not just inference — is the whole point. Pretraining a 1.6T model is the single most Nvidia-dependent workload in AI. It is precisely the thing US export controls on H100/H800-class silicon were designed to make impossible inside China. LongCat-2.0 is the first public claim that it has been done anyway, at trillion-parameter scale, start to finish, on domestic accelerators.

**One caution, because the headlines overreached.** Meituan did *not* name its chip vendor. Its own materials say "AI ASIC superpods" and credit Huawei's HCCL communication library — the domestic analogue to Nvidia's NCCL — for training stability. SCMP noted plainly that *"Meituan did not explicitly name its hardware supplier,"* and that both Meituan and Huawei declined to comment. So the correct claim is not "trained on Huawei Ascend 910s," however widely that inference is drawn. The correct claim is narrower and still remarkable: **a trillion-parameter model was pretrained and served entirely without Nvidia hardware, on Chinese silicon whose vendor the company chose not to confirm.** Meituan's technical report even conceded the constraint — its domestic accelerators have *"significantly less memory per device than Nvidia's H800 chip,"* and *"compared to the mature Nvidia GPU ecosystem, the supporting software community is still less developed."* They shipped anyway.

# The gate and the giveaway are the same story

For two weeks this publication has traced Washington's move to insert itself into the frontier-model release process — the [Anthropic export shutoff](/posts/2026-06-27-government-joined-the-model-release/), the [GPT-5.6 Sol gate](/posts/2026-07-09-the-voluntary-gate-that-works-like-a-license/), the "trusted partner" whitelist. The animating logic of that entire regime is a single assumption: **that the strategic chokepoint is access to the most capable model, and that access can be controlled** — by a letter to a lab, by an export rule on a chip, by a review that decides who gets Sol and when.

LongCat-2.0 is the assumption's stress test, and it fails in two directions at once.

It fails on the model, because you cannot gate an MIT download. There is no trusted-partner list for a file that has already propagated to every mirror on earth. The US regime's core instrument — revocable, curated access — has no purchase on a weight release. A government can stop OpenAI from shipping Sol for twelve days. It cannot un-ship LongCat-2.0 from the machines that already have it.

And it fails on the chips, because the export-control theory assumes the frontier *needs* the silicon the controls restrict. Pretraining a 1.6T model on domestic hardware, whatever the vendor, is a direct claim that the dependency is loosening. If the best models can be trained without Nvidia, the chip chokepoint that underwrites the entire gate — the reason "access to a model" was ever treatable as an "export" — starts to leak.

Put the two fortnights side by side. Washington's answer to frontier risk was to make the release a licensed event. China's answer to the license was to make the release un-licenseable — give the model away, under the most permissive license that exists, trained on chips the export regime cannot reach. These are not two unrelated stories filed under "policy" and "open source." They are the same contest, run from opposite ends. The gate assumes scarcity can be manufactured and controlled. The giveaway assumes it cannot, and moves to make sure it isn't.

# The counterargument, taken seriously

The skeptic has a real case. LongCat-2.0 is not GPT-5.6 Sol. It trails the closed frontier on reasoning, its own benchmark table admits it, and "near-frontier agentic coding model" is a narrower claim than "frontier model." The domestic-chip run came with candid caveats — less memory per device, a thinner software ecosystem — that suggest efficiency and cost penalties the press release doesn't quantify. And "trillion-parameter model trained on domestic hardware" is, for now, Meituan's own assertion, corroborated by reporting but not independently audited at the silicon level. A determined skeptic can read the whole thing as a well-timed marketing claim wrapped around a good-but-not-frontier coding model.

That reading is too comfortable, for the same reason the "it was just eighteen messy days" reading of the Anthropic ban was. It mistakes the current gap for a permanent one. The relevant number in [our June 24 analysis](/posts/2026-06-24-open-weight-reasoning-gap-three-months/) was not that open weights had caught the frontier — they hadn't — but that the *lag had collapsed to roughly a quarter-year* and was still shrinking. LongCat-2.0 doesn't need to beat Sol. It needs to be good enough at the workload buyers actually pay for — agentic coding — while shipping under a license and on a hardware base that the US regime has no lever against. On SWE-bench Pro, it clears that bar today. The question is not whether the giveaway matches the gate on raw capability this week. It is whether "access to the best model" survives as a chokepoint when a good-enough model ships free, worldwide, on chips no one can embargo.

# Why this matters

The instinct is to file LongCat-2.0 under "another strong Chinese open-weight model" and move on. That undersells what the July 4 weekend actually staged: a direct collision between two theories of how model power is controlled.

For US policy, the implication is uncomfortable. The frontier gate is built on chip scarcity and access control, and LongCat-2.0 pressures both — one release claiming Nvidia-free trillion-parameter pretraining, shipped under a license that makes access control moot. Every gated US release now has a shadow: a downloadable alternative that asks no one's permission, improving on a schedule Washington doesn't set.

For the labs, the pricing pressure we flagged in [the Grok 4.5 week](/posts/2026-07-08-three-labs-shipped-one-asked-permission/) now has an open-weight floor under it. A near-frontier agentic-coding model you can self-host under MIT removes both per-token cost and vendor lock-in from the equation — the exact two constraints the closed labs monetize.

And for the strategic picture, the lesson of the fortnight is that the two most important AI-policy events of early July point in opposite directions and belong in the same frame. The US made a model release something you need clearance to do. China made a comparable-enough release something no one can stop. A gate only works if there's a wall around it. The giveaway is a bet that the wall is already full of holes — and that the fastest way to prove it is to hand everyone a copy of what's on the other side.

## Sources

- [Hugging Face — LongCat-2.0 model card (Meituan)](https://huggingface.co/meituan-longcat/LongCat-2.0)
- [GitHub — meituan-longcat/LongCat-2.0](https://github.com/meituan-longcat/LongCat-2.0)
- [South China Morning Post — China debuts biggest AI model trained on local chips as Meituan releases LongCat-2.0 (Jun 30, 2026)](https://www.scmp.com/tech/tech-trends/article/3358854/china-debuts-biggest-ai-model-trained-local-chips-meituan-releases-longcat-20)
- [VentureBeat — Meituan open-sources LongCat-2.0, the 1.6T near-frontier agentic coding model trained entirely on Chinese chips (Jun 29, 2026)](https://venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic-coding-model-thats-been-leading-openrouter-trained-entirely-on-chinese-chips)
- [SiliconANGLE — China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips (Jun 30, 2026)](https://siliconangle.com/2026/06/30/chinas-meituan-open-sources-massive-longcat-2-0-ai-model-saying-trained-domestic-chips/)
- [Caixin Global — Meituan open-sources 1.6-trillion-parameter AI model built on Chinese chips (Jul 7, 2026)](https://www.caixinglobal.com/2026-07-07/meituan-open-sources-16-trillion-parameter-ai-model-built-on-chinese-chips-102461495.html)
- [Open Source For You — Meituan open-sources LongCat-2.0 under MIT license (Jun 30, 2026)](https://www.opensourceforu.com/2026/06/meituan-open-sources-longcat-2-0-under-mit-license/)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-07-04-the-gate-and-the-giveaway/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt