Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 6 min read

LongCat-2.5 is MIT. The download is not a laptop

Meituan put LongCat-2.5 on Hugging Face on November 4, 2026, under MIT. The card says 1.6T total and about 48B active. It is not a laptop download.

RA
· Founder, AIToolRanked
A dim server hall with rows of dark racks and one open tray of chips lit amber in the foreground
On this page

What landed on Hugging Face on November 4

Meituan published LongCat-2.5 at meituan-longcat/LongCat-2.5 ↗. The Hugging Face repository was created at 2026-11-04 11:48 UTC, and the oldest commit on main is an initial commit at that same timestamp. Weight-shard commits follow later the same day. The model card was edited again on November 5, November 9, and November 10, 2026.

That is the weights date. Posts on X on November 11 still describe the drop as new. The Decoder ↗ published on November 15. November 9 is a model-card commit, not the first upload.

The picture above is ours. We have not run the model.

The model card ↗ calls LongCat-2.5 a native multimodal mixture-of-experts model for long-horizon agent work, with 1.6 trillion total parameters and about 48 billion activated per token. The safetensors header on the same repo reports 1,779,752,648,320 parameters, about 1.78 trillion, almost all BF16. Hugging Face's stored size for the repo is 3,561,989,945,589 bytes, about 3.56 TB. GPUJournal ↗, dated November 9, reports 3,562 GB, the same size. Size a disk from the byte count. Quote 1.6T and ~48B active as Meituan's description.

The LICENSE ↗ file is the MIT license, copyright 2026 Meituan. It does not grant Meituan trademarks or patents. Chat is at longcat.ai ↗.

The longcat.ai blog page's description reads "1.6万亿总参大模型,训练全程由国产芯片完成": a 1.6-trillion-parameter model whose training was completed entirely on domestic chips. The card says the work used domestic AI accelerators, and that a W4A4 build targets native INT4 on those accelerators, with NPU notes pointing at SGLang-FluentLLM. A second repo, LongCat-2.5-W4A4 ↗, is also tagged MIT. The Decoder's headline says the model never touched an Nvidia GPU. The sentences we are using are Meituan's: domestic chips, and domestic AI accelerators.

How you try it

The hosted API and the checkpoint are different context numbers.

LongCat's API docs ↗, fetched November 15, list model id LongCat-2.5, an OpenAI-compatible base at https://api.longcat.chat/openai, and an Anthropic-compatible base at https://api.longcat.chat/anthropic. The same docs include setup pages for Claude Code, Codex, OpenCode, and Hermes. Context on that API is 1,048,576 tokens, counting input and output together. The max_tokens cap is 262,144. Thinking is a switch.

The pricing page ↗ lists a limited-time pay-as-you-go rate per million tokens: $0.30 uncached input, $0.006 cached input, $1.20 output. The same page shows ¥2, ¥0.04, and ¥8. It says the discount can change, and that the platform and your bill are the prices that count.

The weights file does not repeat the API window. config.json sets max_position_embeddings to 524,288. Do not quote the API's 1,048,576 as the checkpoint's config.

OpenCode's model docs ↗ list LongCat 2.5 Free (longcat-2.5-free): free for a limited time, 100 requests a day, hosted in China, prompts not used for training and not retained. The free list moves. Check it live.

The card's usage example sends text and an image and asks for text back.

What the card claims on benchmarks

Every number below is from Meituan's card, fetched November 15. A score marked cited is one Meituan says came from a public report. An unmarked score is in-house, on Meituan's definition. We did not run these tests. The Claude column is labeled Claude-Opus-5, not Claude Opus 5.5.

BenchmarkLongCat-2.5Claude-Opus-5 on the card
Terminal Bench v4.038.553.9, cited
Terminal Bench v2.188.489.1, cited
BrowseComp91.990.8, cited
AutomationBench v1.0.651.850.3, cited

The 88.4 figure in circulation is Terminal Bench v2.1. It is not v4.0, where LongCat is 38.5 and the card's Claude-Opus-5 score is 53.9.

Benchmarkkimi-k3 on the cardDeepSeek-V4.1-Flash on the card
Terminal Bench v4.012.6, cited31.2, cited
Terminal Bench v2.188.3, cited90.6, cited
BrowseComp91.2, cited91.5, unmarked
AutomationBench v1.0.646.7, cited54.8, cited

Those headers are the card's labels. They are not a retest of Kimi K3.1 Pro, and they are not the Fireworks speed tier in our DeepSeek V4.1 Flash Ultrafast note.

What people are saying

  • "Open source claims: read the license. Meituan's LongCat-2.5 is a 1.6T MoE under MIT, commercial use included." — Nathan ↗ (@Muskmelon0706), November 14

  • "the part I'd watch is the W4A4 build with NPU deployment notes, because self-hosting on Chinese chips is where a license this loose actually changes decisions." — HUMAN ALGORITHM ↗ (@ztr666ztr), November 11

  • "The test of “open” is now practical: who can run it reliably?" — spaisee ↗ (@spaisee_com), November 11

  • "OpenCode's free roster rotates, so check the live list before you count on it." — Bobo ↗ (@okeyccronald), November 11

  • "the company that brings your lunch now ships frontier open weights." — Lost in AI Translation ↗ (@SiliconIntel_CN), November 11

Reddit was not reachable for this story.

Our take

Read the license, then the disk. MIT, including commercial use, is the loose part. Kimi K3.1 Pro is open weights under Moonshot's license, with a revenue tripwire. LongCat's file is the standard MIT text. About 3.56 TB and a 1.6T-class mixture-of-experts is a cluster download. Edge0-145B is the laptop-class post. llama.cpp 1.0 is the local runner, for a model that fits the machine.

Try the API before you order hardware. The listed limited-time rate, $0.30 in and $1.20 out per million, sits under Claude Opus 5.5 at $4 and $20 and Sonnet 5.5 at $2 and $10 on Anthropic's pricing page ↗. OpenCode says the free row is hosted in China and does not retain prompts. If prompts have to stay on your machines, that row is the wrong place. The MIT weights are the path that matches the license, on a cluster that can hold them.

The card is not a win over Claude. LongCat leads the cited Claude-Opus-5 scores on BrowseComp and AutomationBench, and trails both Terminal Bench versions, by a wide gap on v4.0. The DeepSeek-V4.1-Flash column leads LongCat on AutomationBench and Terminal Bench v2.1. All of those figures are vendor-reported.

Who should care. Agent builders who want a long hosted context at that token price, and shops that can run a huge mixture-of-experts. Laptop shoppers should stay with Edge0. The wider downloadable field is the open-source LLM comparison. Price context for a seat you already pay for is cheaper Claude Code alternatives.

More from the news desk · About Rai Ansar