Field intelligence for AI-first professionalsVol. II · Nº 195 · Monday, November 16, 2026
← Phantom Notes
Nº 171AI ModelsNovember 10, 20264 min read

Meituan Open-Sourced LongCat-2.5 Under MIT. The Only Deployment Guide Is for 194 Huawei Chips, and the API Price Is Still Limited-Time.

Meituan published LongCat-2.5 and a W4A4 quantized build on Hugging Face under the MIT license, repositories created November 4 and last modified November 10, with public release dated November 10 to 11 by posts on X. That resolves the open question from our November 1 post, which found no open weights announced. The README says the model can be deployed on NPU platforms and points to SGLang-FluentLLM, whose guide describes a 194-card Huawei Atlas A2 cluster for the W4A4 build; we found no NVIDIA guidance. Hugging Face metadata totals about 1.78 trillion parameters against the 1.6 trillion Meituan claims, a gap we flag and do not explain. The W4A4 throughput claims and the benchmark table are Meituan's, with rival scores self-reported, and the $0.30 and $1.20 API price is still limited-time with no end date. We read the repos, the NPU guide, and the pricing page.

By T.W. Ghost


Meituan has published the weights for LongCat-2.5. The Hugging Face repositories meituan-longcat/LongCat-2.5 and LongCat-2.5-W4A4 carry the MIT license in both the README and the LICENSE file; they were created on November 4 and last modified on November 10, 2026, at 07:28 UTC, and posts on X date the public release to November 10 and 11. Our scouts missed it, and The Decoder covered it on November 15, so this note is late. It resolves the one open question in our November 1 post, which found a 1.6 trillion parameter mixture of experts, about 48 billion active, with a limited-time API price and no open weights announced. Now there are weights, under the most permissive common license there is.

The catch is where you can run them. The README says the model "can be deployed on NPU platforms" and points to SGLang-FluentLLM, and the deployment guide there describes a Huawei Atlas A2 cluster of 194 cards for the W4A4 build: two for the encoder, 64 for prefill, 128 for decode. We found no NVIDIA GPU guidance in what we read. Meituan's W4A4 numbers, relayed by The Decoder, claim 43.3% more prefill throughput and 42.77% more decode throughput than 8-bit at accuracy "on par"; those are vendor figures we did not recheck in the README. One more oddity: the safetensors metadata on the BF16 repo totals about 1.78 trillion parameters, against the 1.6 trillion Meituan states. That may be a vision tokenizer or extra layers; we flag it and do not assert either. We read both repos, the NPU guide, The Decoder, and the pricing page. Here is what held up.


What Shipped

LongCat-2.5 weights (Hugging Face, modified November 10, 2026)
Reposmeituan-longcat/LongCat-2.5; meituan-longcat/LongCat-2.5-W4A4; created Nov 4; last modified Nov 10, 07:28 UTC
LicenseMIT (README and LICENSE)
Deployment"Can be deployed on NPU platforms"; SGLang-FluentLLM; NPU guide: Huawei Atlas A2, 194 cards (2 encoder, 64 prefill, 128 decode) for W4A4; no NVIDIA guidance found
W4A4 claims+43.3% prefill, +42.77% decode throughput vs 8-bit; accuracy "on par" (vendor, via The Decoder)
Parameter count1.6T claimed; safetensors metadata totals about 1.78T on the BF16 repo; unexplained
README benchmarksTerminal Bench v2.1 88.4 (Opus 5 89.1*, GPT-5.6 Sol 88.8*, DeepSeek V4.1 Flash 90.6*); Terminal-Bench 4.0 38.5 (Opus 5 53.9*); rivals starred as self-reported
API price$0.30 in / $0.006 cached / $1.20 out per 1M; "limited-time," "subject to change," no end date (unchanged)
OpenCodelongcat-2.5-free, 100 requests a day, two weeks free from about Nov 9 (X posts; not checked on OpenCode)

MIT is the headline. No territory clause, no revenue gate, no hosted-service gate, which is more than MiniMax's H3-Image or Kimi's custom license offered this month.

One hundred ninety-four Huawei chips is the fine print. Open weights you can only run, by the only guide published, on an Ascend cluster are open in law and narrow in practice. Someone will write the NVIDIA recipe; Meituan did not.

The benchmark table compares against last generation. Opus 5, GPT-5.6 Sol, and the rivals' own self-reported numbers. On Terminal-Bench 4.0 the README shows 38.5 against Opus 5's 53.9, which is not a line Meituan put in a headline.

Still no price you can plan on. Limited-time, subject to change, and no end date, as it was at launch.


Verified vs Unconfirmed: The Scorecard

ClaimVerdict
Weights on Hugging Face; MIT license; repo datesVerified (README; LICENSE; Hugging Face API)
Public release Nov 10 to 11X posts; repo modification date Nov 10
NPU deployment pointer; 194-card Atlas A2 guide; no NVIDIA guide foundVerified (README; SGLang-FluentLLM guide)
W4A4 throughput and accuracyVendor (via The Decoder; not rechecked in README)
About 1.78T in metadata vs 1.6T claimedVerified discrepancy; unexplained
README benchmark rowsVendor; rival scores self-reported
API price unchanged and limited-timeVerified (pricing page)
OpenCode free tierUnverified (X posts)

Who Should Use It

Use the MIT weights if you have Ascend hardware or the engineering to port a 1.6 trillion parameter MoE to anything else, and you want a license with no strings.

Skip the "never touched an NVIDIA GPU" framing as a feature until there is a guide for the hardware you own, and skip budgeting on the API price until it has an end date.

The frontier is elsewhere: the open-weights leader on Artificial Analysis is Kimi K3.1 Pro at 46, and the open model with the most providers behind it is DeepSeek V4.1 Flash; LongCat-2.5 has no AA row we have read. Our model comparison has the current standings side by side.


Sources


*Not sure whether an MIT license you cannot run on your hardware counts as open for you? Take the free 2-minute quiz and get matched. Then read the launch post this one closes out: LongCat-2.5 at launch.*

Which model should you be using?

Three minutes, twelve questions, one defensible answer.

Take the quiz →