Meituan open-sources LongCat-2.5: a 1600B-parameter multimodal MoE (48B active, 1M context) under MIT license

Meituan open-sources LongCat-2.5: a 1.6 trillion-parameter multimodal MoE (48B active, 1M context) under the MIT license

🔎 A Chinese giant releases 1.6 trillion parameters as open weights

Meituan has just crossed a line that no American lab has dared to cross: publishing the weights of a frontier model with 1.6 trillion parameters under the MIT license. On November 10, 2026, the Chinese delivery and e-commerce giant put the weights of LongCat-2.5 online on Hugging Face — a natively multimodal MoE designed for long agentic tasks. Commercial use, fine-tuning, redistribution: everything is permitted.

Why now? Because the gap between open models and closed models has never been thinner — and because China continues to widen its lead on the open weights front. After DeepSeek and Qwen, after GLM-5.2 and its 753B MoE under the MIT license, Meituan is shifting into a higher gear with a model that accepts images, handles a 1M-token context, and costs only $0.30/M input tokens via its API.

Meanwhile, the American labs — OpenAI, Anthropic, Google — remain firmly closed-weights. The battle for open AI is now being fought almost exclusively on the Asian and European side. LongCat-2.5 is the latest episode, and it is significant.

Key takeaways

• Meituan releases the LongCat-2.5 weights under the MIT license: commercial use, modification, and redistribution allowed without restriction (the license does not cover Meituan trademarks or patents).
• Natively multimodal MoE architecture: 1.6T total parameters, roughly 48B activated per token, thinking and non-thinking modes, image input supported, 1M token context.
• Massive gains vs LongCat-2.0: Terminal Bench v2.1 at 88.4 (vs 70.8), DeepSWE v1.1 at 64.6 (vs 9.7), BrowseComp at 91.9, SWE Atlas at 58.3 (vs 33.1).
• The picture remains mixed against rivals: DeepSeek-V4.1-Flash (90.6) and Claude Opus 5 (89.1) outperform LongCat-2.5 on Terminal Bench v2.1; on Terminal Bench v4.0, LongCat-2.5 (38.5) trails GLM-5.3 (41.8) and Claude Opus 5 (53.9).
• Too big for personal hardware, but 4-bit quantization of expert weights and NPU deployment via an SGLang fork already exist.
• Meituan API at $0.30/M input tokens and $1.20/M output tokens (limited-time offer, November 2026). No third-party inference providers yet.

Recommended tools

LongCat-2.5: what exactly is it? A multimodal MoE designed for agents

Direct answer: LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters, of which about 48 billion are activated per token, with image input, a 1M token context, and two reasoning modes (thinking and non-thinking). It's the first model of this size released under an MIT license with native multimodal capabilities.

The key word here is "natively multimodal". This isn't a text model with a vision module bolted on after the fact. Images are part of the architecture from pre-training onward, which changes the game for agentic tasks: browsing, document reading, interface manipulation. This is exactly the niche Meituan is targeting, as it specifies on Hugging Face that the model is optimized for long agentic tasks.

The MoE architecture with 48B active out of 1.6T total is the classic compromise of the moment: you get the capacity of a colossal model, but the inference cost of a much smaller one. It's the same logic as GLM-5.2 with its 753B MoE, and it's what makes these models economically viable — 1.6T of dense parameters simply couldn't be served at any reasonable price.

The 1M token context is the other serious argument. For agents that need to browse dozens of web pages, entire codebases, or document corpora, it's a prerequisite that has become quasi-standard on frontier models — and LongCat-2.5 offers it with no licensing cost.

Thinking and non-thinking modes

The model offers two generation modes. Thinking mode engages explicit reasoning before answering, useful for complex coding and analysis tasks. Non-thinking mode answers directly, for simple tasks where latency matters. This is a pattern that has become standard, but its implementation in an open weights model of this size remains rare.

Benchmarks: the numbers are impressive, but not dominant

Direct answer: LongCat-2.5 makes a spectacular leap over LongCat-2.0, but it doesn't dominate any leaderboard against the best models of the moment. It's a very high-level model, not a number-one model.

Let's start with the internal progression, which is genuinely spectacular:

Sources: CuriousLM and VentureBeat. The gain on DeepSWE v1.1 — from 9.7 to 64.6 — is the most striking: it suggests the previous generation was practically unusable for autonomous software engineering, and that the new generation has become competitive. This is the kind of jump you see between two major generations, not between a minor iteration.

Now, the table against rivals, and here it's more nuanced:

My reading: on Terminal Bench v2.1, LongCat-2.5 is in the leading pack, 2.2 points behind DeepSeek-V4.1-Flash and 0.7 points behind Claude Opus 5. On Terminal Bench v4.0 — the more recent and more demanding version of the benchmark — the gap widens: 38.5 versus 53.9 for Claude Opus 5. So the model is excellent on benchmarks from its generation, but the hardest version of the test reveals a limitation against the most recent closed models.

This is a pattern I'm seeing more and more: open models crush established benchmarks, but newer, harder benchmarks restore a gap in favor of closed models. No miracle here, no hype: LongCat-2.5 is a very good frontier model, not a model that makes Claude or GPT obsolete.

The MIT license: why it's the real story

Direct answer: the MIT license is the most permissive license that exists, and applying it to a 1.6T-parameter multimodal model is a first. That's what sets LongCat-2.5 apart from most Chinese "open weights" releases, which are often published under custom licenses with usage restrictions.

Concretely, the MIT license allows:

• Commercial use without restriction — you can integrate LongCat-2.5 into a paid product, without negotiation, without revenue share.
• Fine-tuning and modification — you can adapt the model to your domain and publish your variants.
• Redistribution — you can host, resell, and redistribute the weights.

The only exclusions: Meituan's commercial trademarks and patents. VentureBeat confirms that the release is indeed under a permissive license, without restrictive clauses like "acceptable use" or "fewer than 700M monthly active users" that you find elsewhere.

Why does it matter? Compare it with the current alternatives. DeepSeek V3.1 had already moved the needle with its MIT license, but on a smaller model. Aleph Alpha released Kolibri-1, a 78B MoE (3.46B active) under Apache 2.0, focused on on-premise sovereign AI. And Thinking Machines is about to unveil Inkling 2 Small under Apache 2.0. The trend is clear: permissive licenses are becoming the standard on the open weights front, and LongCat-2.5 pushes that standard to maximum scale.

For organizations and cloud providers, this is the decisive factor. An MIT license means a cloud provider can host LongCat-2.5 and offer it to its customers without legal risk. That's exactly what made Llama a commercial success back in the day — and what's sorely missing from models under custom licenses.

Can you run it locally? No, not really — and that's the price of scale

Straight answer: with 1.6T total parameters, LongCat-2.5 is beyond the reach of any personal hardware, even quantized. It's a datacenter model, not something for your RTX 4090.

Meituan implicitly acknowledges this. CuriousLM is blunt: the model is too large for personal hardware. But two deployment paths already exist for the best-equipped:

• A SGLang fork enables deployment on NPUs, a rare option for open weights models of this size.
• 4-bit quantization of expert weights on home accelerators reduces the memory footprint — but even at 4 bits, 1.6T parameters represent hundreds of gigabytes, out of reach of a consumer setup.

Concretely, to run LongCat-2.5 locally, you need a multi-GPU cluster or cloud access. This is the domain of organizations, not individuals.

If you want local, the realistic strategy in 2026 is hybrid: an API for heavy workloads (LongCat-2.5, or an equivalent model), and a smaller local model for everyday tasks. Our selection of the best local LLMs and the guide to installing a local LLM with Ollama or LM Studio cover that part. And if you're building agents, our guide on open source AI agents with Ollama shows how to combine a small local model with API calls for demanding tasks.

No inference provider for now

Important point: as of publication, no third-party inference provider (Together, Fireworks, Groq, etc.) hosts LongCat-2.5. The only turnkey option is Meituan's own API, at $0.30/M input tokens and $1.20/M output tokens — a limited-time offer, clearly designed to accelerate adoption. It's an aggressive price, but the dependence on a single API, hosted in China, is a factor to consider for companies sensitive to data sovereignty. This is precisely the niche where Aleph Alpha positions Kolibri-1 for on-premise sovereign AI.

China Is Widening the Gap on Open Weights — and US Labs Are Looking Elsewhere

Straight answer: nearly all of the frontier open weights models released in 2026 come from China or Europe. The major American labs haven't released anything in open weights, and LongCat-2.5 accentuates that imbalance.

Look at the list of recent open weights releases: DeepSeek, Qwen, GLM from Z.AI, Kimi from Moonshot, LongCat from Meituan. On the Western side, there's Aleph Alpha with Kolibri-1, Thinking Machines with Inkling 2 Small, and a few niche players. None of the big American labs — OpenAI, Anthropic, Google — has released a frontier model in open weights.

Why the gap? Three reasons, in my view:

1. The business model. US labs monetize APIs and subscriptions; releasing weights destroys their moat. Chinese players use open weights as an ecosystem weapon: they gain adoption, community contributions, and influence in areas where they can't win on API revenue.
2. Regulatory pressure. Europe and parts of Asia are pushing for auditable, on-premise deployable models. Open models are the only ones that meet this need.
3. The network effect. Every MIT-licensed model released attracts developers, fine-tunes, and tooling. DeepSeek figured this out, Z.AI figured it out, Meituan figured it out.

The practical consequence for you: if you're building on open models, your choice is almost exclusively among Chinese or European models. Our monthly comparison of the best LLMs reflects this imbalance, and closed models still lead on the most recent benchmarks — but the gap is narrowing month after month.

Where does LongCat-2.5 stand in this hierarchy?

In the open weights leading pack, alongside DeepSeek-V4.1-Flash and GLM-5.3. On coding benchmarks, DeepSeek-V4.1-Flash keeps a slight edge on Terminal Bench v2.1, and GLM-5.3 outperforms LongCat-2.5 on Terminal Bench v4.0. But LongCat-2.5 brings a unique combination: native multimodality + 1M context + MIT license + 1.6T size. No other open model checks all four boxes simultaneously.

How to use LongCat-2.5 concretely: three scenarios

Direct answer: three access paths exist — the Meituan API, cloud self-hosting, and downloading the weights for fine-tuning. The choice depends on your volume and your sovereignty constraints.

Scenario 1: testing via the API

The simplest option. The Meituan API costs $0.30/M input tokens and $1.20/M output (limited-time offer, November 2026, check on tech.meituan.com). For comparison, that's an aggressive pricing position for a model of this class. The protocol is standard, so integration into an existing pipeline is quick. If you host your application on a VPS, Hostinger works very well for the front end — the Meituan API handles inference.

Scenario 2: cloud self-hosting

For organizations that want to control inference without buying GPUs. Plan for a serious multi-GPU cluster — the model is too big for personal hardware, even quantized. The SGLang fork with NPU support is the path documented by Meituan, with 4-bit quantization of the expert weights to reduce the footprint. This is an infrastructure project, not a weekend deployment.

Scenario 3: fine-tuning and redistribution

This is where the MIT license shines. You can download the weights from Hugging Face, fine-tune on your domain, and publish your variant — commercially, without restriction. For cloud providers, it's a direct opportunity: becoming the first third-party inference provider for LongCat-2.5, since CuriousLM confirms that none exists to date. The first one to jump in captures the demand.

Should You Adopt LongCat-2.5? My No-Nonsense Verdict

Direct answer: yes for organizations that want a multimodal open-weights frontier model with an MIT license; no for individuals and small projects, which have neither the hardware nor the need.

Adopt it if you're an organization that:

• needs a multimodal agentic model with a large context window, without depending on a closed lab;
• wants fine-tuning on your data with an unrestricted license;
• is a cloud provider looking for a new model to host;
• does research and wants an auditable frontier model.

Move along if you're:

• a solo developer or a small team: the Meituan API is an option for testing, but the dependence on a single Chinese provider and the absence of third-party inference providers make the continuity risk real. For day-to-day coding, our comparison of the best LLMs for coding remains more relevant.
• someone looking for purely local solutions: the model is too big. Instead, check out our selection of the best local LLMs and the installation guide.
• someone looking for a free model for personal use: the options in our selection of the best free LLMs are better suited.

My take: LongCat-2.5 isn't the model that will change your daily life as a developer this week. It's a model that changes the landscape over twelve months. An MIT license at this scale, with native multimodality and a 1M context, sets a precedent that the other players — Chinese and Western alike — will have to follow. And for organizations building agents for the long term, it's exactly the kind of foundation they're looking for: our guide to the best LLMs for AI agents should, incidentally, include LongCat-2.5 in its next update.

❌ Common Mistakes

Mistake 1: Confusing "open weights" with "open source"

The weights are released under the MIT license, which is indeed open source in the strict sense. But beware: the training data, the pre-training code, and the full recipes are not published. LongCat-2.5 is an open weights model under a permissive license, not a fully open source project in the manner of Linux. The distinction matters when you're evaluating reproducibility.

Mistake 2: Believing you can run it on your own PC

1.6T parameters — even with 48B active and 4-bit quantization of the experts, that's still hundreds of GB of memory. The SGLang fork for NPUs and the 4-bit quantization target specialized setups, not a gaming PC. The solution: use the API for heavy tasks, and a small local model for the rest.

Mistake 3: Trusting benchmarks from the previous generation

The gains vs LongCat-2.0 are spectacular (+54.9 points on DeepSWE v1.1), but that's not the right comparison. The right comparison is against today's rivals: DeepSeek-V4.1-Flash and Claude Opus 5 outperform LongCat-2.5 on Terminal Bench v2.1, and the gap widens on Terminal Bench v4.0. Always judge a model against its current competitors, not against its own previous version.

Mistake 4: Ignoring the absence of third-party inference providers

No third-party provider hosts LongCat-2.5 at this time. If you're building a product on the Meituan API, you have a single point of failure, hosted in China, with "limited-time" pricing. Have a plan B ready — or wait for a cloud provider to announce hosting, which shouldn't take long given the MIT license.

❓ Frequently Asked Questions

Is LongCat-2.5 really free?

The weights are free under the MIT license: commercial use, fine-tuning, and redistribution are permitted. The only exceptions are Meituan's trademarks and patents. Meituan's API, on the other hand, is paid ($0.30/M input tokens, $1.20/M output tokens, limited-time offer in November 2026). Downloading the weights is free, inference is not — unless you have your own infrastructure.

Can you run LongCat-2.5 on a Mac or a gaming PC?

No. With 1.6T total parameters, the model far exceeds the capabilities of any consumer hardware, even when quantized to 4 bits. Meituan offers an SGLang fork for NPU deployment and a 4-bit quantization of the expert weights, but these options target specialized accelerators. For local use, pick a smaller model from our selection of the best local LLMs.

Is LongCat-2.5 better than Claude Opus 5?

No, not on the most recent benchmarks. On Terminal Bench v2.1, Claude Opus 5 (89.1) edges out LongCat-2.5 (88.4). On the more demanding Terminal Bench v4.0, Claude Opus 5 (53.9) clearly dominates LongCat-2.5 (38.5). LongCat-2.5's appeal isn't about beating Claude, but about offering a comparable level with downloadable weights and an MIT license.

Why is Meituan releasing a 1.6T-parameter model?

Meituan is a giant of Chinese e-commerce and delivery, with massive needs for agents (customer support, logistics, browsing). Releasing the weights costs it little — it keeps the infrastructure — and pays off in adoption, community contributions, and influence over the ecosystem. It's the same strategy as DeepSeek and Z.AI: winning the open weights game, where US labs defend their API revenue.

How reliable are Meituan's benchmarks?

The benchmarks are self-reported by Meituan, which calls for standard caution. VentureBeat confirms the gains vs LongCat-2.0, and the figures against rivals are published alongside competitors' results — including when LongCat-2.5 loses, which is a sign of good faith. Expect independent evaluations in the weeks following the release.

✅ Conclusion

LongCat-2.5 is not the most powerful model in the world — but it is the most powerful multimodal model ever released under an MIT license, and that changes the game for organizations that want to build on open foundations. China continues to widen the gap on open weights; the question is no longer whether US labs will follow, but when. To track how the landscape evolves month by month, check out our comparison of the best LLMs and our selection of the best local LLMs.