Meituan open-sources LongCat-2.5: 1.6T params, MIT licence

Meituan spent the autumn building a 1.6-trillion-parameter model and selling it through an API. On Monday night it gave the weights away instead.

Meituan's LongCat team released the full weights of LongCat-2.5 under the MIT licence, making a frontier-class multimodal model free to download, fine-tune and resell — no usage caps, no revenue thresholds, no request for permission. The model went live as a hosted API on November 2; what changed overnight is that anyone with the hardware can now run it themselves. Alongside the main checkpoint, Meituan published a W4A4 quantised build and deployment notes for domestic Chinese accelerators, per the model card on Hugging Face.

LongCat-2.5 is a native multimodal mixture-of-experts model: 1.6 trillion total parameters, roughly 48 billion activated per token, and a 1-million-token context window. Meituan built it for long-horizon agentic work rather than chat — complex codebase development, full-project builds, visual programming that produces interactive 3D scenes, and business analysis spreadsheets. The model card lists the design bets plainly: native-resolution images converted into discrete tokens and trained jointly with text, long-horizon reinforcement learning on real engineering tasks and reproduced agent failures, and multi-teacher distillation from specialised agent, reasoning and interaction models.

The licence is the part worth pausing on. Chinese labs have been the most prolific open-weight publishers of the past year, but most of those releases ship under bespoke licences with attribution clauses, user caps or commercial restrictions. MIT is the most permissive licence in common use: commercial deployment, redistribution and further training all come free. Meituan's card says it in one line — "The model weights are released under the MIT License" — and adds that nothing in the card should be read as restricting those terms.

The benchmark table is impressive and narrower than the headline number suggests. Across 31 evaluations and nine comparison models — including Claude Opus 5, GPT-5.6-Sol, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash — LongCat-2.5 finishes first on four: BrowseComp at 91.9, ALE v1.1 at 38.5, ResearchRubrics at 81.0 and 3DCodeBench at 55.4. Two of those wins are photo finishes. BrowseComp is 91.9 against DeepSeek's 91.5 and Kimi K3's 91.2; ResearchRubrics is 81.0 against DeepSeek's 80.9. The other two are real gaps: ALE puts it 3.3 points clear of Kimi and 8.2 ahead of GPT-5.5, and 3DCodeBench — generating 3D code from a text prompt — is 1.1 points ahead of Claude Opus 5 and the only category where a Chinese model leads at all.

The losses are just as instructive. On Terminal Bench v4.0, the hardest agentic terminal test in the table, LongCat-2.5 scores 38.5 against Claude Opus 5's 53.9 — a 15-point gap that no amount of BrowseComp leadership offsets. It trails DeepSeek on DeepSWE (64.6 versus 74.2), on AutomationBench (51.8 versus 54.8) and on JobBench (65.4 versus Claude Opus 5's 66.7). On Terminal Bench v2.1 the two leaders are effectively tied, 88.4 to 89.1, which makes the v4.0 spread the more interesting number: the new benchmark separates models that the old one could not.

Read the footnotes before quoting any of it. Meituan marks scores taken from other labs' published reports with an asterisk and states that unmarked scores were "measured in-house and are indicative only." That is honest labelling, and it also means the four first-place finishes are a vendor's own measurements on a vendor's own harness. Treat them as a claim to be reproduced, not a settled ranking.

Where LongCat-2.5 is unambiguously competitive is price. The hosted API is running a limited-time rate of 2 yuan per million input tokens and 8 yuan per million output tokens — around 28 and 113 US cents — with cached input at 0.04 yuan. Zhidx reported that this undercuts the four overseas models in the comparison table by a wide margin, and that the gap narrows against domestic rivals: DeepSeek-V4.1-Flash's updated checkpoint charges the same peak rate and half of it off-peak, and beats LongCat-2.5 on 14 of the 26 tests where both report scores. Against DeepSeek's own 1.6-trillion-parameter V4-Pro, LongCat-2.5 is cheaper on all three unit prices.

Adoption in the first hours was fast enough to be a data point of its own. OpenCode, the open-source coding agent, put the model in front of its users free for two weeks at 100 requests a day, and by noon Hong Kong time on Tuesday LongCat-2.5 had logged about 109 billion tokens, 4,700 unique users and roughly 50,000 sessions on the platform. That is a fraction of what a ChatGPT-scale model sees in a minute, but for a model released as weights less than twelve hours earlier it is a real signal that developers were waiting for exactly this: a big Chinese model they can point their existing Claude Code, Codex or OpenCode setup at.

The engineering notes are aimed squarely at the constraint Western labs do not have. Meituan trained and serves LongCat-2.5 on domestic AI accelerators with under 80 GB of memory per card, sharding parameters and expert exchange across supernodes and overlapping communication with compute; its long-sequence fine-tuning configuration reports 86% higher per-device throughput than a standard distributed baseline. The W4A4 build quantises expert weights and activations to four bits and, against an INT8 baseline, lifts prefill throughput 43.3% and decode throughput 42.77% while cutting time-to-first-token 15.9% and time-per-output-token 19.7%. Squeezing a 1.6T model onto export-controlled silicon is the strategic problem Chinese labs are actually solving, and this release documents the solution in public.

Two caveats keep this from being a clean sweep. Developers on X are already reporting looping behaviour — the model re-derives the same conclusion and burns tokens doing it — which matters more for an agentic model that runs unattended than the same flaw would on a chatbot. And the practical barrier is unchanged by a permissive licence: the checkpoint is a 1.8-terabyte safetensors repository, so "free to download" and "free to run" remain very different sentences. Hugging Face's download counter showed 195 for the last month at the time of writing, which tells you the early audience is hosting providers and research groups, not hobbyists.

The strategic read is that this is a distribution play dressed as a gift. Meituan is a food-delivery and local-services company; it does not sell compute at scale, and it gains nothing from developers self-hosting except mindshare and a halo around its API. Giving away MIT-licensed weights buys something an API cannot: a place in the default toolchain. Meituan shipped setup instructions for Claude Code, Codex, OpenCode and Hermes on day one, which is an argument about where developers already live, not about benchmark scores. It follows the pattern set by Alibaba's Qwen3.8-2.4T-A95B open release and Tencent's Hunyuan Hy4 preview — three Chinese labs in three months shipping trillion-scale weights while the American frontier stays closed.

What to watch: independent evaluations. Artificial Analysis and LMArena will tell us within a week whether 91.9 on BrowseComp survives a harness nobody at Meituan controls, and whether the Terminal Bench v4.0 gap is a training-data artefact or a real ceiling on long-horizon execution. The other thing to watch is whether MIT holds if the model gets popular — a licence is only as durable as the company's willingness to keep honouring it at scale.

Is a 1.6-trillion-parameter model you can download but barely serve actually open, or is the cheap API the real product? Write to us via our contact page.

Arete-27B: Nous cut Qwen's propaganda rate from 22% to 0.9%

Nous Research took Alibaba's Qwen3.8-27B, put about fifty million tokens of supervision through it, and shipped back a model that answers political questions like a Western lab's build. The interesting part is how little it took — and that the whole thing is checkable. Nous Research has released Arete-27B, an Apache 2.0 model post-trained from Alibaba's Qwen3.8-27B, and says the share of answers containing Chinese state propaganda fell from 22 percent to 0.9 percent while general capability…

Mistral unveils Le Chonk: a 1T-parameter open-weights model

The biggest open-weight release outside China lands in public preview today, and the country that spent the week promising its own frontier model just put a price on the ambition. Mistral has opened a public preview of Mistral Large 4 — codenamed "le Chonk" — a 1 trillion-parameter mixture-of-experts model with 49 billion active parameters, natively multimodal, which the company calls its largest and most capable model to date.

Qwen open-sources a 7B image model that writes real transparency

Qwen cut its image model down to a size consumer cards can hold and gave it something no other open generator ships natively: an alpha channel. In Washington, the Congressional Black Caucus says the votes are there for AI guardrails. Alibaba's Qwen team open-sourced Qwen-Image-2.1, a 7B image model that generates transparent images natively — real RGBA layers instead of a white box you mask out afterward.

Tencent open-sources Hunyuan Hy4 preview — 770B params, 1M context

China's open-weight race just got its heaviest new contender, and Tencent is pitching it as a productivity workhorse rather than a benchmark-chaser. Tencent is open-sourcing Hy4 preview, a 770 billion-parameter Mixture-of-Experts model with a 1 million-token context window. The architecture activates only 49 billion parameters per token — a 78-layer stack where each layer beyond the first holds 256 routed experts plus one shared expert, with eight experts firing per token.

Qwen3.8-27B hits 1M downloads as small open models close the frontier gap

Alibaba's latest open-source release is turning heads: Qwen3.8-27B, a 27-billion-parameter model that runs on personal computers, has crossed one million downloads within days of launch and is outperforming models several times its size on independent benchmarks. The model scored 52 on the Artificial Analysis Intelligence Index — placing it above every medium-class open model (40B–150B parameters) and tying with DeepSeek V4 Flash, which ranks fifth among all large models.

Zhipu's GLM-5.3 ships with 'emergent' cyber capabilities

China's open-weights race just produced its most awkward win: Zhipu's GLM-5.3 landed Thursday with frontier-level coding and a cyber capability its own makers admit outgrew their expectations — and Moonshot AI is racing a closing window toward a Hong Kong listing. Zhipu released GLM-5.3 on Thursday, an open-weights flagship whose every gain comes from post-training — including an "emergent" cyber capability the company says developed faster than it expected.

DeepSeek's V4 Pro open weights finally land on Hugging Face

DeepSeek finished the job it started two days ago: the flagship's open weights are out, IBM bet its consulting army on OpenAI's stack, and the AI data center boom headed for the public market. DeepSeek has released the open weights for V4 Pro — build 0813, the production flagship it shipped this week — on Hugging Face, under an MIT license, roughly 890 GB of weights that anyone can now download and self-host.

Qwen open-sources Qwen3.8-2.4T-A95B, its first Max-class model

Wednesday's wire: Alibaba opened the weights on its most powerful model ever, Honor put a robotic arm in a phone and sold 400,000 of them before lunch, and Europe's booksellers started refusing suspicious bulk orders. Alibaba's Qwen team open-released Qwen3.8-2.4T-A95B on Wednesday — the first Qwen-Max-class model to ship as open weights, with 2.4 trillion total parameters and 95 billion active in a 512-expert mixture-of-experts architecture.

Get SiliconNoon in your inbox

Read more