On 10 November 2026 Meituan published the LongCat-2.5 open weights on Hugging Face under the MIT licence: a 1.6 trillion parameter mixture-of-experts model with about 48 billion parameters active per token, in a 3.56 TB repository. On licence terms it beats Kimi K3.1 Pro and Baidu's ERNIE 5.5 Flash Base, because neither revenue gates nor territory carve-outs apply. For most Indian developers and teams the sensible order is to rent the hosted API first, at roughly ₹28 per million input tokens (pricing), and self-host only if you already operate trillion-parameter MoE serving.
TL;DR
- Meituan released the weights one week after launching LongCat-2.5 as a hosted service on 1 November 2026.
- The licence is plain MIT. It grants no rights to Meituan trademarks or patents, which matters for legal review.
- The checkpoint is 3.56 TB across 208 safetensors shards. Meituan's tested serving path targets NPUs, not GPUs.
- Hosted pricing is a limited-time discount: $0.30 input and $1.20 output per million tokens, about ₹28 and ₹114.
- Benchmarks are mostly in-house and flagged by Meituan as indicative only.
- Verdict: rent first, run later. MIT makes it the lowest-risk large open-weights base to build on this month.
What changed from LongCat-2.0?
LongCat-2.5 builds on LongCat-2.0 and adds native visual input alongside agentic coding. According to the model card, vision comes from LongCat-Next, which turns native-resolution visual features into discrete tokens aligned with text. It is a hybrid model: thinking and non-thinking modes are switched per request with chat_template_kwargs: {enable_thinking: ...} in the OpenAI-compatible API.
| Benchmark (model card, in-house) | LongCat-2.0 | LongCat-2.5 |
|---|---|---|
| Terminal Bench v2.1 | 70.8 | 88.4 |
| Terminal Bench v4.0 | 3.0 | 38.5 |
| DeepSWE v1.1 | 9.7 | 64.6 |
| BrowseComp | 79.9 | 91.9 |
| AutomationBench v1.0.6 | 15.8 | 51.8 |
Meituan states that unmarked scores are measured in-house and are indicative only, so treat the jumps as a direction rather than a guarantee. On the same card, Terminal Bench v2.1 puts Claude Opus 5 at 89.1 and DeepSeek-V4.1-Flash at 90.6, both cited from public reports, so LongCat-2.5 is close to the leaders rather than ahead. VentureBeat notes that the comparison set predates newer models, some of which post higher Terminal Bench v4.0 scores than LongCat's 38.5; even Claude Opus 5's cited 53.9 on that suite is well clear of the 38.5, per the model card comparison.
How does the MIT licence compare with Kimi K3.1 Pro and ERNIE?
The large open-weights releases this November now differ more by licence than by capability. The LongCat LICENSE file is 1.06 kB and reads "Copyright (c) 2026 Meituan", with the model card confirming the weights are MIT.
| Model | Licence | Main constraint |
|---|---|---|
| LongCat-2.5 (Meituan) | MIT | None beyond attribution; no trademark or patent grant |
| Kimi K3.1 Pro (Moonshot) | Modified MIT | Separate agreement above a revenue threshold for model-as-a-service use; see our Kimi K3.1 Pro write-up |
| ERNIE 5.5 Flash Base (Baidu) | Custom | Excludes the EU, UK and South Korea; commercial use needs notice or authorisation by revenue band |
The Kimi and ERNIE rows come from third-party write-ups of their licence files, so read the originals before committing a product to either. If you are choosing a base model for a company, MIT removes a whole category of legal questions. For a comparison of two other recent releases, see Tencent HY4 vs Mistral Large 4.
How do I download the LongCat-2.5 open weights?
The repository shows this command pattern:
hf download hf://meituan-longcat/LongCat-2.5
Expect 208 shards of roughly 17.2 GB each, plus an 8.12 MB index file. Check free disk before starting; the checkpoint alone is 3.56 TB, before any converted or quantised copies. The upgrade path from LongCat-2.0 is a straightforward swap of model path and chat-template flags, provided your serving stack supports the architecture.
That is the catch. Hugging Face lists no inference provider hosting the model, and Meituan's documented serving route is NPU platforms through SGLang-FluentLLM. GPU users should wait for community vLLM or SGLang support. Meituan reports W4A4 quantisation on expert weights, with 43.3% higher prefill throughput and 42.77% higher decode throughput than a W8A8 baseline, per the model card. Those figures describe Meituan's own stack, not what you will see on other hardware.
Should I rent the API or run LongCat-2.5 myself?
Rent first. The LongCat pricing page lists a limited-time discount of $0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens. Using the RBI reference rate of 94.86 rupees per dollar from 3 November 2026, that is approximately ₹28, ₹0.57 and ₹114. All rupee figures here are approximate and will move with the exchange rate and the end of the discount.
A worked example: at the listed promo rates of $0.30 per million uncached input tokens and $1.20 per million output tokens (LongCat pricing), an agent pipeline using 10 million uncached input tokens and 2 million output tokens a day comes to $3.00 plus $2.40, about $5.40 a day, roughly ₹512 at those rates (pricing page), around ₹15,400 over 30 days. That is far below the cost of keeping a multi-node cluster busy for a 3.56 TB model.
| Situation | Better choice |
|---|---|
| Prototype, evaluation, spiky traffic | Hosted API |
| Existing trillion-parameter MoE serving, NPU access | Self-host |
| Data must not leave your network | Self-host, once GPU support matures |
| Need a legally clean base to fine-tune or redistribute | Download the MIT weights |
The hosted price is promotional, so rerun the numbers when it ends. Also note the model is listed on Vercel's AI Gateway, per TLDR, which gives a second route to try it.
Why does this release matter for developers searching for open models?
Specific model names are where developers actually search, and they are a narrow target. We priced 694 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only 72 (10.4%) cleared a winnable bar of 150-6,000 monthly searches, difficulty 20 or below, a genuine technical term, and at least three words. (n=694, measured 2026-11-09, keyword volume and difficulty data via DataForSEO). A concrete query such as "LongCat-2.5 open weights" is the kind of term that passes that bar, because it names a real artefact. If you are weighing coding-agent tooling to pair with it, our guide to open-source agent platforms for coding is a reasonable next read.
FAQ
Q: Is LongCat-2.5 really open source?
A: The weights are published under the MIT licence, which is permissive. Training data and full training code are not part of the release, so "open weights" is the more accurate term.
Q: Can I use LongCat-2.5 commercially in India?
A: The MIT licence places no revenue or territory limits on commercial use. It does not grant rights to Meituan trademarks or patents, so do not use Meituan branding without permission.
Q: Can I run LongCat-2.5 on GPUs today?
A: Meituan's tested path is NPU-based through SGLang-FluentLLM. No public inference provider hosts it yet, so GPU users should wait for community adapter support.
Q: How big is the LongCat-2.5 download?
A: The repository is 3.56 TB across 208 safetensors shards of about 17.2 GB each, according to the Hugging Face page.
Q: Are the LongCat-2.5 benchmark scores reliable?
A: Treat them as indicative. Meituan labels unmarked scores as in-house measurements, and the comparison models are not the latest releases.
Q: Are smaller open models worth considering instead?
A: For narrow tasks, often yes. A specialised model such as PLaMo 3 Translate runs offline and fits on far more modest hardware, and Arete 27B shows what a mid-sized post-train can do.
Top comments (0)