Meituan LongCat-2.5 open weights: 1.6T-parameter MoE under MIT, 88.4 on Terminal Bench v2.1
Meituan released LongCat-2.5 on Hugging Face, a multimodal agentic model with 48B active parameters; its own tests put it near the top on Terminal Bench but mid-pack on DeepSWE.
· 6 min read
Written by AI from the sources linked below. How we use AI
On this page
Meituan, the Chinese delivery and local-services company, has put the weights of LongCat-2.5 on Hugging Face under the MIT licence. The model is a multimodal mixture-of-experts system with 1.6 trillion parameters in total and about 48 billion active per token, aimed at long agentic tasks such as working through a codebase. On Meituan's own tests it lands within a point or two of the best closed and open coding agents on one headline benchmark and well behind on another. The catch for most readers is hardware: the only deployment guide on the model card targets Chinese NPUs, not the GPUs most developers rent.
Key points
- LongCat-2.5 is released as open weights under the MIT licence at meituan-longcat/LongCat-2.5 on Hugging Face (the repository was created on 4 November 2026 and last updated on 10 November), with a 4-bit variant, LongCat-2.5-W4A4.
- Meituan's model card gives 1.6 trillion total parameters and roughly 48 billion activated per token; Hugging Face's own metadata counts about 1.78 trillion stored parameters, a gap the card does not explain.
- On Meituan's in-house run, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, but 64.6 on DeepSWE v1.1, below several rivals' reported scores.
- The model card's only deployment guide is for NPU platforms through SGLang-FluentLLM; it gives no GPU recipe.
- VentureBeat reports a limited-time hosted API price of $0.30 per million uncached input tokens and $1.20 per million output tokens.
What happened
According to the model card, LongCat-2.5 builds on LongCat-2.0 and adds native visual understanding to agentic coding, with the stated goals of complex codebase development, full-project builds, creative design and 3D tasks. VentureBeat, which reported the weights release on 9 November, says Meituan introduced the model about a week earlier and then published the weights on Hugging Face. The repository's metadata on Hugging Face shows it was created on 4 November and last modified on 10 November at 07:28 UTC, within seconds of its 4-bit sibling.
Meituan says the vision side works by converting native-resolution image features into discrete tokens that are aligned with text in one shared space, an approach it took from its earlier LongCat-Next work. For training it describes long-horizon agentic reinforcement learning built on real engineering tasks and reproducible agent failures, with safeguards against reward hacking, plus "multi-teacher" learning in which separate agent, reasoning and interaction teachers guide the model. It cites an 86% gain in per-device throughput over a DDP baseline in reported long-sequence supervised fine-tuning, and says training and inference logic are aligned so token log probabilities match bit for bit across training, prefill and decoding. These are the company's claims from its own model card; we could not test them.
Background
LongCat is Meituan's model family. Its predecessor LongCat-2.0 also had 1.6 trillion total parameters and was published on Hugging Face in early July 2026, according to the repository's modification dates; we did not confirm earlier details of its launch. LongCat-2.5 therefore reads as an upgrade of an already open giant, not a new architecture from scratch.
Open weights at this scale matter because few teams can run them, but many can study, distil or fine-tune from them, and an MIT licence permits commercial use, modification and redistribution as long as the licence notice stays. For another recent open release, see our report on Nous Research's Arete 27B; and for the closed frontier it is being compared against, Anthropic's Claude Fable 5.2 and Mythos 5.2.
The numbers
Meituan's model card compares LongCat-2.5 with LongCat-2.0 and eight other models. Scores marked with an asterisk in the card are cited from other labs' reports or public leaderboards; the rest were measured in-house and the card calls them "indicative only". Several rivals' figures are therefore not run under the same setup, so treat the tables as a guide and not a head-to-head.
On this test LongCat-2.5 sits in a tight cluster at the top, about 2 points behind DeepSeek-V4.1-Flash (90.6) and under 1 point behind Claude Opus 5 (89.1) and GPT-5.6-Sol (88.8). The jump from LongCat-2.0 is 17.6 points. The card says its Terminal Bench v2.1 result is averaged over three runs with a Claude Code harness.
DeepSWE tells a more modest story: 64.6, seventh of nine, about 10 points behind the top score. Compared with LongCat-2.0's 9.7 the gain is large, but against peers the model is mid-pack. On the harder Terminal Bench v4.0 it scores 38.5 against 53.9 for Claude Opus 5 and 41.8 for GLM-5.3 (LongCat-2.0 scored 3.0); VentureBeat notes that some leaderboard comparisons were run under different setups. Other results from the card: 91.9 on BrowseComp, 83.1 on MCP Atlas, 78.6 on Vision2Web (Claude Opus 5 leads at 80.2) and 55.4 on 3DCodeBench, the highest in its table.
How it works
LongCat-2.5 is a mixture-of-experts model: for each token a router picks a small set of experts, so only about 48 billion of 1.6 trillion parameters do work per token. Compute per token is therefore closer to a mid-sized dense model, but all the weights must still be held in memory, which is the real barrier to running it.
The model card describes a W4A4 scheme for the expert weights and activations, using FlatQuant-style learnable transformations to keep accuracy comparable to the unquantized model on the benchmarks it evaluated. Against a W8A8 mixture-of-experts baseline, it reports in online stress tests 43.3% higher prefill throughput, 42.77% higher decode throughput, 15.91% lower time to first token and 19.72% lower time per output token; with KV-cache quantization (W4A4KV8) decode throughput is 92% higher than W8A8KV16. These numbers are Meituan's, measured on its own serving stack, and the card says the INT4 computation is native on "domestic AI accelerators".
On size, Hugging Face's API lists the full model at roughly 1.78 trillion parameters (mostly BF16) and the W4A4 repository at about 495 billion stored elements in packed integer types. The full repository's reported storage is about 3.6 TB and the 4-bit variant's about 1.3 TB, although that storage figure may include more than one copy of files, so treat it as an upper bound. VentureBeat's illustrative figure of roughly 3.2 TB for 1.6 trillion parameters at 16-bit is consistent with the order of magnitude. Either way, this is a multi-node job, not something for a single workstation.
What it means
For developers: the model is free to study and build on, but self-hosting is a serious infrastructure project. The model card says LongCat-2.5 "can be deployed on NPU platforms" and points to SGLang-FluentLLM; no vLLM or GPU instructions appear in the card we read. Teams on other hardware would need to validate a serving stack themselves. The practical route for most is the hosted API; VentureBeat reports limited-time pricing of $0.30 per million uncached input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens, and we have not checked this against Meituan's price page.
For the open-weight ecosystem: another trillion-parameter model under an MIT licence raises the floor for what open weights can do on terminal and web-agent tasks, and gives researchers a base for distillation into smaller models.
For hardware watchers: Meituan frames the training and serving optimizations around domestic Chinese accelerators. That is a company statement about its stack; it says nothing about the chips' relative speed.
What we don't know yet
- The exact parameter count: 1.6 trillion on the card versus about 1.78 trillion in the stored tensors.
- A clear statement of context length. The card's footnotes mention evaluation sequence lengths of 256K, 512K and 1M depending on the test, but we did not find one headline context window.
- Independent results. All LongCat-2.5 scores we cite are Meituan's own, and some rival numbers come from other sources under different setups.
- Safety evaluations and behaviour outside benchmarks, and the memory needed to serve it on GPUs.
What to watch next
Look for third-party runs on public leaderboards, a community GPU serving recipe or GGUF-style conversions, and Meituan's technical blog and pricing page as the limited-time API price changes. Also watch whether the 4-bit W4A4 variant, which needs native INT4 support, runs well outside the NPUs it was built for.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
PyTorch 2.15: rewritten NCCL backend is now default, Python 3.10 dropped
The release makes the new NCCL backend the default for GPU training, adds a hang-time flight recorder, and raises the minimum Python and macOS versions.
audiotale.net team7 min read
Tencent Hy4 launches Nov 4: weights due 'in the coming days', API price and tests
Tencent's full Hy4 adds vision and computer use and is on sale by API now, but only the August preview weights are on Hugging Face so far.
audiotale.net team7 min read
Ripple-65B Preview: 65B open-weight coding model, 80.5 on TerminalBench 2.1, Apache 2.0
Ricursive Intelligence says its dense 65B model beats US open models up to 15 times larger on agentic coding for $32 million; the benchmarks are self-reported.
audiotale.net team7 min read