LIVE TICKER
[Live wire]Connecting to the newsroom…
TechnologyPublished 11 November 2026 • 10 min read • 2,017 words • 6 cited sources

LongCat-2.5 open source: Meituan publishes 1.6T weights under MIT

Meituan has published the LongCat-2.5 weights on Hugging Face under the MIT licence. The agentic model claims 1.6 trillion parameters and 48 billion active per token, but the file metadata adds up to about 1.78 trillion.

LongCat-2.5 open source: Meituan publishes 1.6T weights under MIT
Meituan LongCat / Hugging Face — view at source ↗

Key facts at a glance

Total parameters
1.6 trillion claimed; about 1.78 trillion in Hugging Face metadataMeituan model card / Hugging Face
Active parameters
About 48 billion per tokenMeituan model card
Licence
MIT for the model weightsHugging Face model card
Checkpoint size
3.56 TB across 208 safetensors shardsHugging Face
API price
$0.30 input / $1.20 output per million tokens, limited-timeLongCat API Platform
Terminal Bench v2.1
88.4, measured in-house by MeituanMeituan model card

What did Meituan release, and when?

Meituan has published the full LongCat-2.5 weights on Hugging Face under the MIT licence. The Chinese food-delivery and e-commerce company introduced the model on 2 November. TechNode reported at the time that GitHub and Hugging Face releases were planned to follow.

The repository now holds 208 safetensors shards. Its commit history shows the shards uploaded on 5 November, hours after the repository was created, with model-card revisions on 9 and 10 November. AI Weekly dates the public release to 9 November.

The launch made LongCat-2.5 available through Meituan's chat service and API platform. The weights matter more to researchers, because they allow inspection, fine-tuning and self-hosting without Meituan's permission.

Meituan is not a conventional AI lab. Its business is delivery, local services and e-commerce, and its LongCat team has existed since 2023, according to The Decoder. The team shipped its first model in late 2025.

Publishing a trillion-parameter checkpoint under a permissive licence is a strategic move. It invites outside developers to build on Meituan's model rather than a rival's, and it generates attention that a closed API cannot.

The release also lands during a busy stretch for Chinese open weights. AI Weekly, which tracks the sector, says LongCat-2.5 is the latest of 304 China AI stories it has logged in ninety days, many of them open-weight releases.

The release followed a preview cycle. Meituan's API changelog lists LongCat-2.5-Preview on 25 September and the full LongCat-2.5 on 2 November, with enterprise billing added earlier in September.

How large is LongCat-2.5 really?

Meituan's model card states 1.6 trillion total parameters, with about 48 billion active for each token. That gap is normal for a mixture-of-experts design, which sends each token through a small group of specialists rather than the whole network.

Hugging Face's metadata gives a different total. The safetensors index adds up to 1,779,752,648,320 parameters, or roughly 1.78 trillion. That is about 11% above the figure in Meituan's announcement.

AI Weekly flagged the same mismatch and warned that it makes hardware sizing unreliable for teams planning to self-host. The index splits the count into 1.78 trillion BF16 parameters and 1.23 billion F32 parameters.

The counting gap may come from what is included. Parameter totals can cover embeddings, the vision tower, the visual quantiser or only the language experts, and Meituan does not publish a breakdown.

The download is not modest. Hugging Face reports 3.56 TB of stored files, with the weights split across 208 shards. A full checkpoint needs serious storage and a serving plan before anyone benchmarks it.

At the time of writing, the model page showed 195 downloads and 81 likes. Those are small numbers for a release this large, which suggests most interest is still exploratory.

The configuration file fills in the architecture. It declares 38 layers, 768 routed experts plus 128 identity experts, and 12 experts selected per token. The vocabulary is 163,840 tokens.

The context window is 524,288 positions, or 512K tokens. Attention uses multi-head latent attention with an index-based sparse selection that keeps 2,048 tokens, plus 1,024 local tokens.

The configuration declares 524,288 positions, which is 512K tokens. Meituan's evaluation notes set harness sequence limits of 512K for most coding tests and 1M for DeepSWE, so the usable window depends on how the model is served.

Active parameters explain the cost profile. Routing each token through 48 billion parameters keeps inference far cheaper than a dense model of the same total size, which is why labs use this design for high-volume agent work.

The visual side uses a 32-layer vision transformer with 14-pixel patches. A residual quantiser with eight codebooks of 16,384 entries turns image features into discrete tokens, which are aligned with text in a shared space.

Training used multi-teacher learning, with specialised agent, reasoning and interaction teachers guiding the model through full-vocabulary distillation. Meituan also reports 86% higher per-device throughput than a DDP baseline on long-sequence fine-tuning.

Reinforcement learning targeted long-horizon work. Meituan says it turned real engineering tasks and reproducible agent failures into training signal, with safeguards against reward hacking.

Meituan applies four-bit quantisation to expert weights and activations, a technique it calls W4A4. The card claims accuracy comparable to the unquantised model on the benchmarks it evaluated.

Which benchmarks does LongCat-2.5 win?

Meituan's evaluation table covers code-agent, general-agent and visual tasks. The gains over LongCat-2.0 are large almost everywhere.

Terminal Bench v2.1 rises from 70.8 to 88.4. DeepSWE v1.1 jumps from 9.7 to 64.6, SWE Atlas codebase questions from 33.1 to 58.3, and Program Bench from 2.5 to 25.5. AutomationBench climbs from 15.8 to 51.8.

Against rival models the picture is tighter. The 88.4 on Terminal Bench v2.1 sits within a point of GPT-5.6-Sol at 88.8 and Claude-Opus-5 at 89.1, and behind DeepSeek-V4.1-Flash at 90.6.

Harder tests expose bigger gaps. On Terminal Bench v4.0, LongCat-2.5 scores 38.5 against 53.9 for Claude-Opus-5. On DeepSWE it scores 64.6 against 74.2 for DeepSeek-V4.1-Flash and 72.7 for GPT-5.6-Sol.

General-agent results are competitive. BrowseComp reaches 91.9, ahead of Claude-Opus-5 at 90.8 and GPT-5.6-Sol at 90.4. HLE with tools reaches 63.1, behind Claude-Opus-5 at 64.7 but ahead of GPT-5.6-Sol at 58.0.

Other agent tests are less flattering. On SkillsBench it scores 63.3 against 69.8 for DeepSeek-V4.1-Flash. On AutomationBench it scores 51.8 against 54.8 for the same model. GDPVal-AA rubrics put it at 84.8, behind 87.7 for both Claude-Opus-5 and GLM-5.3.

Office-style tasks show the same pattern. MCP Atlas reaches 83.1 against 85.8 for Claude-Opus-5, JobBench reaches 65.4 against 66.7, and SpreadsheetBench 2 reaches 34.9 against 35.8 for GLM-5.3.

Two more results round out the coding picture. PaperBench reaches 82.1, ahead of 79.1 for Claude-Opus-5 and behind 83.3 for GPT-5.6-Sol. Claw-Eval reaches 78.3 and LongCatClawBench reaches 83.3, behind 86.3 for Kimi K3.

Visual results are mixed. MathVision reaches 96.5 and CharXiv-RQ 90.2, but VideoMME lands at 85.2 against 89.5 for GPT-5.6-Sol. Vision2Web reaches 78.6 against 80.2 for Claude-Opus-5. Text-to-3D is a bright spot at 55.4, the best score in its table.

The methodology deserves attention. Meituan marks rival scores with an asterisk when they come from public reports, and calls its own unmarked results in-house and indicative only. Harnesses range from Claude Code 2.1.280 to mini-swe-agent and Codex.

Several results are averaged over three runs, and judges include Claude Opus 4.5 and gemini-3.6-flash. Those choices are disclosed, which is better than silence, but they still make cross-model comparison approximate.

AI Weekly noted that no third-party benchmark validation existed at launch. That does not make the numbers wrong. It does mean buyers should treat them as vendor claims until independent runs appear.

Was LongCat-2.5 trained on Chinese chips?

Meituan's card credits infrastructure optimisations on domestic AI accelerators for the training run. It does not name the supplier, and the weights alone cannot prove which chips were used.

The predecessor offers context. The Decoder reported that LongCat-2.0 was trained on more than 50,000 domestically made AI ASICs and over 35 trillion tokens.

Meituan said at the time that LongCat-2.0 proved it could train large models on domestic computing clusters. The company did not identify the chipmaker, which left independent verification difficult.

LongCat-2.5 extends that engineering work with four-bit expert quantisation. Meituan reports 43.3% higher prefill throughput and 42.77% higher decode throughput than an eight-bit baseline, with 15.91% lower time to first token.

It also reports a 1.73-times speed-up from grouped matrix multiplication kernels. Those are internal stress-test figures rather than independent audits.

Serving gets a similar treatment. Meituan says a four-bit weight and cache configuration delivers 92% higher decode throughput than an eight-bit baseline, and that its cache optimisation nearly doubles decode capacity on the tested workload.

The inference stack is part of the pitch. Meituan says its aligned training and serving logic produces bitwise-identical token probabilities across training, prefill and decoding, a claim about reproducibility rather than raw speed.

The card also credits communication and computation overlap for long-context training efficiency. It gives no wall-clock training time, which makes the efficiency claims hard to compare with other labs.

The context is export control. The Decoder notes that US restrictions have applied since 2022, limiting Chinese access to advanced Nvidia accelerators. Domestic silicon is therefore a strategic necessity as much as a technical choice.

If Meituan's claims hold, they support the argument that controls slow but do not stop trillion-parameter training in China. They do not show that domestic chips match Nvidia on cost, yield or reliability.

Meituan has published no such comparison, and the model card gives no training duration, cluster size or failure rate. Those omissions are the difference between a capability claim and a reproducible result.

What does the MIT licence let you do?

MIT is one of the most permissive software licences. It allows commercial use, modification and redistribution, provided the copyright and permission notices are kept.

Meituan's card adds a limit. The licence does not grant rights to Meituan trademarks or patents, which matters for anyone planning to brand a product around the model.

Open weights are not the same as open source. MIT covers the checkpoint, not the training data or the training pipeline, and Meituan has disclosed neither.

Redistribution has obligations. Anyone shipping a derivative must include the MIT notice, and Meituan's card asks developers to assess accuracy, safety and fairness before use in sensitive settings.

Hosted access is cheap during the launch period. The LongCat API lists $0.30 per million uncached input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens. Meituan labels these limited-time prices.

The Chinese price list matches: two yuan per million uncached input tokens, 0.04 yuan for cached input and eight yuan for output. Meituan says platform prices and billing records are authoritative.

At $0.30 per million input tokens, the hosted route is cheap to test. Agent workloads loop through tools and retries, though, so a single task can still consume millions of tokens.

Self-hosting is a different proposition. The checkpoint is 3.56 TB, the configuration loads custom model code, and the deployment guide targets neural processing units through SGLang-FluentLLM.

There is no mainstream GPU recipe in the model card. That is unusual for open-weight releases, which often ship vLLM or SGLang instructions for Nvidia hardware.

Fine-tuning is technically allowed but practically heavy. A full fine-tune needs the checkpoint plus optimiser state; most teams will use adapters or the hosted API instead.

Tooling compatibility is broader than the hardware story. Meituan's API documents guides for Claude Code, Codex, Hermes, OpenCode, Cline and other agent harnesses.

Procurement is the open question. AI Weekly warned that suppliers should check contract terms before using the model for government-facing work, given US scrutiny of Chinese AI.

Open weights make experimentation easy and compliance harder. A team can download the checkpoint without a contract, which also means no vendor indemnity and no data-processing terms.

What should you watch next?

The first thing to watch is independent evaluation. A vendor table with in-house scores is a starting point, not a verdict.

The gap between 88.4 and 38.5 on two terminal benchmarks shows how much test choice matters. Expect third-party leaderboards to publish their own runs within weeks.

The second is documentation. Meituan has not explained the gap between its 1.6 trillion headline and Hugging Face's 1.78 trillion count. A corrected technical report would settle sizing and training questions.

Watch whether Meituan publishes training code. TechNode reported that GitHub and Hugging Face releases were planned; the model card currently links the weights, a technical blog and an NPU deployment guide, but no training repository.

The third is hardware disclosure. Naming the accelerator supplier would let outsiders judge how much of the result depends on domestic silicon rather than model design.

The fourth is price. The $0.30 input rate is explicitly limited-time, and a permanent increase would change the cost comparison with DeepSeek and Qwen.

Watch the licence text as well. MIT is permissive, but the patent and trademark carve-outs in Meituan's card are the clauses legal teams will read first.

The fifth is adoption. Watch for the first Western enterprise to deploy LongCat-2.5 in production and for the contract language it uses.

That will show whether MIT-licensed Chinese weights can clear procurement reviews, or whether they remain a research tool outside China.

For now, LongCat-2.5 is a credible agentic coding model with unusually open terms and unusually thin independent evidence. Both facts belong in any assessment.

Books behind the context

  • Kai-Fu Lee

    Frames the US-China competition over AI talent, data and deployment that makes a Chinese open-weight release a strategic signal.

  • Chris Miller

    Explains the export-control regime and semiconductor supply chain behind Meituan's claims about domestic accelerators.

  • Cade Metz

    Traces the rivalry between open and closed AI labs, the backdrop for judging what an MIT-licensed frontier checkpoint changes.

Sources used for this dispatch

  1. Hugging Face - meituan-longcat/LongCat-2.5 model card
  2. LongCat API Platform - LongCat-2.5 pricing
  3. LongCat API Platform - Change log
  4. TechNode - Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks
  5. AI Weekly - Meituan Open-Sources LongCat-2.5, 1.6T MoE, Under MIT License
  6. The Decoder - Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia

This article is original analysis and synthesis. Facts, figures and quotations are attributed to the sources above. Images remain the property of their publishers and are credited in the caption.

Questions readers ask

Is LongCat-2.5 open source?

Meituan published the model weights under the MIT licence on Hugging Face. The checkpoint is open, but training data and training code have not been released, so it is open-weight rather than fully open source.

How many parameters does LongCat-2.5 have?

Meituan says 1.6 trillion in total, with about 48 billion active per token. Hugging Face's safetensors metadata adds up to roughly 1.78 trillion, and Meituan has not explained the difference.

Can I run LongCat-2.5 on a single server?

Not practically. The checkpoint is 3.56 TB across 208 shards and needs a multi-node serving setup. Meituan's deployment guide targets neural processing units rather than standard GPU servers.

How much does the LongCat-2.5 API cost?

The LongCat platform lists $0.30 per million uncached input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens. Meituan describes these as limited-time prices.

Does LongCat-2.5 beat GPT-5.6 and Claude Opus 5?

It depends on the test. It is within a point on Terminal Bench v2.1 but about 15 points behind Claude Opus 5 on Terminal Bench v4.0. Most of Meituan's scores are in-house rather than independently verified.

What licence does LongCat-2.5 use?

The weights use the MIT licence, which permits commercial use and modification if the copyright and permission notices are retained. Meituan's trademark and patent rights are excluded.

Topics:MeituanLongCat-2.5open-weight modelsmixture-of-expertsChina AIMIT licence

More from the research desk