Subscribe
0%
Illustration for: Meituan opens LongCat-2.5's weights under MIT: a 1.6T vision-and-coding model. What the benchmarks show and what self-hosting takes

Meituan opens LongCat-2.5's weights under MIT: a 1.6T vision-and-coding model. What the benchmarks show and what self-hosting takes

A week after launching it as an API, Meituan put LongCat-2.5 on Hugging Face under the MIT license. Its own numbers put it near GPT-5.6 Sol and Claude Opus 5 on some agent tests and well behind on others, and running it yourself is a data-centre job tuned for Chinese accelerators.

Listen to this article0:00 / 4:48

Meituan has published the weights of LongCat-2.5, its natively multimodal agent and coding model, on Hugging Face under the MIT license. VentureBeat reports the weights went up on November 9, one week after the Chinese delivery giant introduced the model through its chat app and API on November 2. The model card describes a mixture-of-experts (MoE) design with about 48 billion active parameters per token out of roughly 1.6 trillion in total.

Why it matters: MIT is about as permissive as model licenses get. It allows commercial use, modification and redistribution with a notice, which is looser than the custom licenses attached to many "open" models. LongCat-2.5 is also one of the few open-weight models that takes screenshots and video as input and was built for long, multi-step coding jobs, the kind of work most teams currently send to closed models.

What's actually on Hugging Face

Item What the pages say
License MIT for the weights; the card says it grants no rights to Meituan trademarks or patents
Total parameters 1.6T in the card's introduction, but 1.8T in the page header and Safetensors metadata (unexplained)
Active parameters About 48B per token
Inputs Text, images and video; visual input is turned into discrete tokens alongside text
Modes Thinking and non-thinking (enable_thinking)
Context Not given as a spec; evaluation settings mention 256K, 512K and 1M sequence lengths
Variants Meituan's organization page also lists a LongCat-2.5-W4A4 repo (495B in its metadata)
Hosted providers None yet on Hugging Face's inference providers

The 1.6T-versus-1.8T gap was flagged by VentureBeat and Superpower Daily too. It matters if you are sizing hardware, so check the actual checkpoint files rather than the headline number.

The benchmarks, Meituan's numbers

Meituan's model card compares LongCat-2.5 with its predecessor and with several rivals. Starred figures in the card are cited from public sources; the rest are Meituan's own in-house runs, which it calls indicative only.

Benchmark LongCat-2.5 LongCat-2.0 GPT-5.6 Sol Claude Opus 5 DeepSeek-V4.1-Flash
Terminal Bench v2.1 88.4 70.8 88.8 89.1 90.6
Terminal Bench v4.0 38.5 3.0 37.3 53.9 31.2
DeepSWE v1.1 64.6 9.7 72.7 68.8 74.2
SWE Atlas codebase QnA 58.3 33.1 46.0 63.2 55.2
BrowseComp 91.9 79.9 90.4 90.8 91.5
HLE (with tools) 63.1 37.5 58.0 64.7 63.9
Vision2Web 78.6 — 73.5 80.2 79.2

Read plainly: it is a big jump over LongCat-2.0, roughly level with GPT-5.6 Sol and Claude Opus 5 on the older Terminal Bench and on web browsing, and clearly behind on DeepSWE and Terminal Bench v4.0. VentureBeat points out that the rivals Meituan picked are not the newest models; newer systems such as GPT-6.1 Sol and Claude Opus 5.5 post much higher Terminal Bench v4.0 leaderboard scores, though with their own agent harnesses, so the comparison isn't clean. VentureBeat's read is that this places LongCat-2.5 below the frontier. No independent evaluation has been published yet.

What self-hosting takes

This is not a model for a workstation. VentureBeat and Superpower Daily both do the same back-of-envelope sum: at 16-bit precision, 1.6 trillion parameters would be about 3.2TB of weights before any memory for context. Only 48B parameters are active per token, but the whole checkpoint still has to sit in memory.

Meituan's own deployment path is narrow. The model card points to an NPU branch of its SGLang-FluentLLM serving stack and gives no GPU or VRAM figures. The card's W4A4 section says 4-bit expert weights and activations raised prefill and decode throughput by about 43% versus a W8A8 baseline, which describes Meituan's tested setup on Chinese accelerators rather than a guarantee elsewhere.

For most teams the realistic route is the hosted API. VentureBeat lists limited-time launch pricing of $0.30 per million uncached input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens. That is roughly DeepSeek's peak rate and a small fraction of what Anthropic and OpenAI charge for their top models. VentureBeat also notes a cheaper token price doesn't always mean a cheaper finished task if the model needs more retries.

Who should care

  • Teams doing UI and front-end work: VentureBeat's strongest case is jobs that need both visual judgment and code, like turning screenshots into working pages. LongCat-2.5 led Meituan's comparison on 3DCodeBench and sat slightly behind Opus 5 and DeepSeek-V4.1-Flash on Vision2Web.
  • Labs and inference providers: an MIT-licensed, vision-capable model at this scale is a candidate to fine-tune and host. Watch for providers adding it and for GPU serving recipes.
  • Anyone in regulated or government-facing work: VentureBeat found no US federal restriction naming Meituan or LongCat-2.5 as of November 9. An October DHS order targets DeepSeek, which shows how quickly that can change, so check contracts and data-retention terms before sending sensitive code to the hosted API.

Sources

  1. Hugging Face — meituan-longcat/LongCat-2.5 model card huggingface.co
  2. Hugging Face — meituan-longcat organization page huggingface.co
  3. VentureBeat — Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT License venturebeat.com
  4. Dataconomy — Meituan Releases LongCat-2.5 AI Model Weights Under MIT License dataconomy.com
  5. Superpower Daily — Meituan publishes LongCat-2.5 weights, opening a self-hosted route for visual coding superpowerdaily.com
  6. RuntimeWire — Meituan launches LongCat-2.5, a native multimodal model for long-horizon agents runtimewire.com

#Meituan#LongCat-2.5#open weights#MoE#multimodal#agentic coding#Hugging Face#China AI