Meituan Gives Away LongCat-2.5, a 1.6T-Parameter Agent Model, Under MIT Terms
Meituan open-sourced LongCat-2.5, a 1.6-trillion-parameter multimodal MoE trained on domestic Chinese chips, under the MIT license with no revenue thresholds. The food-delivery giant is betting that permissive weights, not a token business, buy it a place in the developer toolchain.
A Food-Delivery Company Just Open-Sourced a Trillion-Parameter Model
Most companies that spend months building a 1.6-trillion-parameter model keep it behind an API. Meituan, China's food-delivery and local-services giant, did the opposite. On November 9-10 its LongCat team published the full weights of LongCat-2.5 on Hugging Face under the MIT license. That is the most permissive license in common use. It allows commercial inference, redistribution and further training with no extra authorization, according to Meituan's own announcement↗.
The model first appeared as a hosted product on November 2. According to Meituan's launch coverage↗, it is a native multimodal model for long-horizon agentic work, trained end to end on domestic Chinese compute. The open-weight drop a week later turns a product launch into an ecosystem play, and it lands at a moment when Chinese labs are racing each other on openness as much as on benchmarks.
For global developers, the headline is simple: a frontier-adjacent, multimodal, 1M-context agent model that anyone can legally download, fine-tune and resell. The harder question is whether anyone can practically run it. This article covers what Meituan shipped, what the numbers say, and where the caveats sit.
What Meituan Actually Released
According to the Hugging Face model card↗, LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion activated per token. It supports a context window of one million tokens. It is positioned for complex codebase development, full-project builds, visual coding, 3D tasks and professional office work.
Key specifications
- Architecture: a native multimodal MoE with 1.6T total and ~48B active parameters, so per-token compute is closer to a mid-sized dense model than to a trillion-parameter one.
- License: MIT, which Meituan says covers commercial inference deployment and secondary training without additional authorization.
- Context: 1M tokens, with long-context training sequences of up to 1M tokens during mid-training according to Meituan's launch materials.
- Quantized build: a W4A4 variant that compresses both expert weights and activations to 4 bits, with deployment guidance via SGLang-FluentLLM for NPU platforms.
- Access: weights on Hugging Face, a technical blog↗, and a hosted API on the LongCat platform.
One discrepancy is worth flagging. AI Weekly reported↗ that Hugging Face metadata at one point listed 1.8T parameters, against the announced 1.6T. The 1.8TB size of the checkpoint on disk may explain the confusion, since bytes and parameters are easy to mix up. We use Meituan's 1.6T figure, which the model card repeats.
How it was trained
Meituan's technical write-up says the long-horizon gains came from a chain of changes, not a single trick. The announcement describes several of them:
- Failure-driven data: real agent failure trajectories are converted into reproducible, verifiable training tasks, so the model learns from the mistakes agents actually make.
- Asynchronous reinforcement learning: partial-rollout training lets unfinished long tasks be saved and resumed, with oversampling adjusted to pass rates and gradient correction for stale trajectories.
- Anti-shortcut constraints: training environments are cut off from the network, git history and file caches are cleared, and an LLM judge supervises trajectories so the model cannot cheat its way to a reward.
- Unified visual tokens: visual signals become multi-level discrete tokens processed by the same backbone as text, rather than through a bolted-on vision module.
"LongCat-2.5 is open-sourced under the MIT license, supporting commercial inference deployment and secondary training with no additional authorization required." — Meituan LongCat team
The Benchmark Picture: Strong, Self-Reported, Uneven
The model card reports results across coding, general agent and visual tasks. Highlights include 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1, 91.9 on BrowseComp, 83.1 on MCP Atlas, 96.5 on MathVision with tools, 85.2 on VideoMME and 78.6 on Vision2Web.
SiliconNoon's analysis↗ is a useful corrective to the headline. It counts first place on four of 31 evaluations in a table of nine comparison models, which it says include Claude Opus 5, GPT-5.6-Sol, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash. Those four are BrowseComp (91.9), ALE v1.1 (38.5), ResearchRubrics (81.0) and 3DCodeBench (55.4). Two of those wins are razor-thin. Elsewhere the model trails.
Where it leads and where it lags
- Leads: web-browsing research (BrowseComp 91.9, narrowly ahead of DeepSeek and Kimi K3 by SiliconNoon's reading) and 3D code generation, where it sits 1.1 points ahead of Claude Opus 5.
- Roughly ties: Terminal Bench v2.1, where it lands within about one point of the top closed models.
- Lags: the harder Terminal Bench v4.0, where it scores 38.5 against Claude Opus 5's 53.9, and DeepSWE, where DeepSeek-V4.1-Flash reaches 74.2 against LongCat's 64.6.
The generational jump over Meituan's own previous release is large. AI Weekly, citing VentureBeat↗, reports Terminal Bench v2.1 rising from 70.8 to 88.4 and DeepSWE from 9.7 to 64.6.
Treat every figure here as a vendor claim. Meituan marks scores copied from other labs' reports and says unmarked numbers were measured in-house and are indicative only. Independent confirmation from evaluators such as Artificial Analysis or LMArena has not been established in the sources we reviewed.
"Meituan's own launch material places LongCat-2.5 between Claude Opus 4.8 and Claude Opus 5 on several mainstream benchmarks, and ahead of Opus 5 on some." — summary of Meituan's launch coverage
The Real Story: Frontier Weights on Constrained Chips
The most strategically interesting part of the release is not the benchmark table. It is the infrastructure disclosure. LongCat-2.5 was trained and served on domestic accelerators where single-card memory is under 80 GB, a hard constraint for long-sequence work. Meituan's response, laid out in its announcement↗, combines several techniques:
- Sharding across supernodes: double-sharded FSDP with flattened buffers and hybrid vocabulary parallelism spread memory pressure across cards.
- Topology-aware communication: parameter all-gathers, expert token exchanges and KV all-gathers are kept inside the high-bandwidth interconnect domain of a supernode wherever possible.
- Dynamic expert replication: MoonEP only switches on when load imbalance crosses a threshold, so extra communication never outweighs the gain.
- Overlap tricks: zero-copy communication, parameter prefetching and ScMoE let communication and compute run concurrently.
The reported payoff is an 86% higher per-device throughput than a DDP baseline on long-sequence supervised fine-tuning. On the inference side, Meituan says its W4A4 scheme, built on FlatQuant-style learnable transforms, improves input-processing and per-token throughput by about 40% against an INT8 baseline. It also cuts time to first token by about 16% and per-token latency by about 20%, with downstream task quality kept comparable on the tasks evaluated.
For the global picture, this matters because Chinese labs are solving a problem their Western counterparts mostly do not have: how to build frontier systems on export-constrained silicon. Publishing the recipe, rather than just the weights, gives the community a reproducible path. It also tells Beijing's chip ecosystem that trillion-parameter MoE work is feasible on local hardware.
Why Give It Away? A Distribution Play
Meituan is not a cloud vendor. On the company's Q2 earnings call, CEO Wang Xing said, as relayed in the launch coverage↗, that the company would not become a "token factory". Its models and AI products are meant to support the core business and improve user, merchant and operating efficiency. LongCat powers an AI operations workbench for local-life merchants, which Meituan says already serves chains such as Guming, Xu Sheng Ji and Yihetang.
That framing explains the license. Meituan does not need API margins. It needs developers, agent frameworks and hosting providers to treat LongCat as a default option, and an MIT license removes every procurement objection except geopolitics. SiliconNoon reports↗ that setup instructions for Claude Code, Codex, OpenCode and Hermes shipped on day one, which suggests the target is the existing developer workflow rather than a new one.
Competitive context
- Versus bespoke licenses: many large Chinese open releases carry revenue thresholds or regional clauses. MIT removes both, which makes LongCat-2.5 easier for startups and enterprises to adopt.
- Versus other Chinese flagships: the same week brought Baidu's ERNIE 5.5 Flash Preview, whose weights are promised later, and Tencent's Hy4. Meituan is, for now, the one that has already shipped weights at trillion scale.
- Versus closed US models: the benchmark claims put LongCat-2.5 near, but not clearly above, the top closed systems on agentic coding, with a visible gap on the hardest terminal tasks.
Caveats Developers Should Weigh
Practical friction is real, and early coverage is candid about it.
- Serving cost: the checkpoint is roughly 1.8 TB, so "free to download" is not "free to run". The early audience is likely hosting providers and research labs rather than hobbyists.
- Hardware guidance: quantization and deployment notes target Chinese accelerators first, which adds friction for teams on H100 or A100 clusters, as AI Weekly noted.
- Behavior reports: SiliconNoon cites early developer reports of looping behavior, where the model re-derives the same conclusion and burns tokens. That matters more for unattended agents than for chat.
- Pricing confusion: launch-period API pricing differs between outlets, so check the official API platform↗ before budgeting. We do not quote a figure here because the reports we reviewed disagree.
- Procurement risk: AI Weekly, relaying VentureBeat, warns that suppliers should check contract terms before using Chinese models in government-facing work.
What to Watch Next
The near-term test is independent evaluation. If third-party harnesses reproduce the BrowseComp and Terminal Bench v2.1 numbers, LongCat-2.5 becomes a credible self-hostable alternative to closed agent models. If the v4.0 gap persists, it will read as a ceiling on long-horizon execution rather than a quirk of training data.
The second test is license durability. MIT is irrevocable for released copies, but a company's willingness to keep releasing future versions on the same terms is a different matter. Meituan's earlier LongCat-2.0 pushed trillion-parameter training onto domestic chips, and 2.5 pushes the weights into the open. A third release in that line, with comparable terms, would show this is strategy rather than a one-off gesture.
The takeaway: LongCat-2.5 is less a benchmark story than an openness story. A company whose core business is food delivery has put a frontier-adjacent, multimodal, million-token agent model into the public domain of commercial use. For builders, the practical move is to wait for independent evals, test the quantized build on your own hardware budget, and keep an eye on looping behavior before putting it near anything unattended. Start with the model card↗ and the technical blog↗.
Links & Resources
External links — opens in a new tab

Scientific Calculators: Treatises and Manuals
by Richard Murdoch Montgomery
The definitive 15-volume series bridging user manuals and applied mathematics — from the TI-Nspire CX II CAS to financial solvers.

Calculus I
by Richard Murdoch Montgomery
Limits, derivatives, integrals, and series — a first course in calculus with formal proofs, worked examples, and applications to physics and engineering.

Artificial Intelligence: Origins and Developments
by Richard Murdoch Montgomery
A comprehensive survey of AI from Turing machines to deep learning — neural networks, expert systems, and the philosophical debates that shaped the field.

The HP 19BII Scientific Financial Calculator
by Richard Murdoch Montgomery
Financial and mathematical reasoning with the HP 19BII — annuities, bonds, cash flows, Solver equations, and regression analysis.
Comments
Open discussion — no account needed. Be respectful.
More from Chinese Models Desk
DeepSeek Ships Its Whole App as an API: Search, Files, Voice and Mandatory Safety Checks on a $0.30 Flash Model
On November 13, DeepSeek exposed the features behind its consumer app through its API: server-side web search, file parsing, URL reading, OCR, ASR and TTS, all anchored on deepseek-flash. The price list and the compliance clause are the real story.
Baidu's ERNIE 5.5 Flash Preview Bets on 14B Active Parameters, With Apache 2.0 Weights Still a Promise
Baidu's 280B-total, 14B-active ERNIE 5.5 Flash arrives as an open base model and a Qianfan-hosted Preview, with a 1M-token context and Apache 2.0 weights reportedly due November 25. We separate what Baidu has documented from what is still a claim.
Alibaba Opens Qwen4-Max Preview: 1M Tokens, Two Weeks, and No Benchmarks
Alibaba's Qwen4-Max preview offers a 1M-token window and video input through Token Plan until November 25, but ships without benchmarks, pricing, weights or a license.