On November 9, 2026, Chinese platform company Meituan released the weights of LongCat-2.5, one week after introducing the model, according to VentureBeat. The checkpoint is on Hugging Face under the MIT License, which permits commercial use, modification and redistribution subject to notice requirements. It is a mixture-of-experts model with about 48 billion active parameters per token out of 1.6 trillion total, and it accepts screenshots, video and codebases as input.
Open-weight releases from Chinese labs have become a regular feature of the model market, and each one raises the same three questions: how good is it, what does it cost to run, and can an organization use it. This post works through the evidence on each, using the figures VentureBeat compiled from Meituan's materials.
What changed from the previous version
Meituan's own comparison against its predecessor shows the largest gains on multi-step software tasks.
| Benchmark | LongCat-2.5 | Predecessor |
|---|---|---|
| Terminal Bench v2.1 | 88.4 | 70.8 |
| DeepSWE v1.1 | 64.6 | 9.7 |
| SWE Atlas codebase QA | 58.3 | 33.1 |
The DeepSWE jump, from 9.7 to 64.6, is more than six-fold and should be treated with care. A very low prior score can reflect a model that could barely attempt the task format, so large relative gains from a low base are common and do not by themselves indicate parity with leaders.
Against rivals: close on some tests, behind on others
VentureBeat's table sets LongCat-2.5 against DeepSeek-V4.1-Flash, GPT-5.6 Sol and Claude Opus 5.
LongCat-2.5 against three rivals on Meituan's selected benchmarks
Scores as reported in VentureBeat's table of Meituan's results, published 9 November 2026, per VentureBeat.
Vendor-reported. Meituan chose older proprietary baselines (GPT-5.6 Sol, Claude Opus 5); newer successors were already available at launch.
The reading depends on the column:
- Terminal Bench v2.1. LongCat-2.5 scores 88.4, within about one point of GPT-5.6 Sol (88.8) and Opus 5 (89.1), and 2.2 points below DeepSeek-V4.1-Flash (90.6).
- DeepSWE v1.1. It trails clearly: 64.6 against 74.2 for DeepSeek, 72.7 for GPT-5.6 Sol and 68.8 for Opus 5.
- Vision2Web and 3DCodeBench. On 3DCodeBench it leads by a narrow margin (55.4 against 54.3 for Opus 5), and on Vision2Web it is 1.6 points behind Opus 5 (78.6 against 80.2).
- Terminal Bench v4.0. Opus 5 scores 53.9 against 38.5 for LongCat-2.5, a gap that the v2.1 results do not show.
Two caveats shape the interpretation. First, Meituan chose its baselines: VentureBeat notes the proprietary models in the table are older, and GPT-6.1 Sol (launched September 29) and Claude Opus 5.5 (announced September 22) were already available. Second, the leaderboard scores VentureBeat lists for those newer models use maximum effort with Codex and Claude Code, so they are not a clean head-to-head. The fair summary is that LongCat-2.5 appears competitive with the previous generation of proprietary models on several coding tests, not that it matches the current frontier.
DeepSeek
OpenAI
AnthropicThe cost picture
VentureBeat calculates the cost of a sample workload of 200,000 uncached input tokens and 100,000 output tokens. At LongCat-2.5's limited-time hosted rate of $0.30 input and $1.20 output per million tokens, that is $0.18.
Cost of one sample workload, by model
VentureBeat's calculation for 200,000 uncached input tokens and 100,000 output tokens, per VentureBeat.
List prices only, in US dollars. Real costs vary with reasoning tokens, repeated tool calls and retries. LongCat-2.5's rate is a limited-time price.
On list prices, the workload costs about 15 times less on LongCat-2.5 than on Claude Opus 5.5 ($2.80) and about 31 times less than on Claude Fable 5.2 ($5.60), by simple division of VentureBeat's figures. DeepSeek's off-peak rate is lower still at $0.09. Two points limit how far that comparison goes. The LongCat rate is described as limited-time, and VentureBeat cautions that real costs vary with reasoning tokens, repeated tool calls and retries. A cheaper model that needs more attempts per accepted result can narrow or erase the gap, which is why cost per accepted result is a better measure than cost per token.
Price per token is a rate. What a team actually pays is price per accepted result, including retries and reviewer time.
On comparing model costs
Self-hosting is a different question
MIT-licensed weights make self-hosting legally possible, but not necessarily practical. VentureBeat notes a discrepancy: Meituan describes 1.6 trillion parameters while Hugging Face metadata shows 1.8T, and it advises sizing deployments from the actual files. As an illustration only, a 16-bit representation of 1.6 trillion parameters would need about 3.2 TB for the weights alone, since each parameter takes two bytes. Meituan's blog describes reducing expert weights and activations to four bits on Chinese accelerators and points to a deployment guide for neural processing units, which suggests the intended path is quantized serving on specific hardware rather than a general-purpose GPU box.
The procurement question
Licensing and procurement policy are separate. VentureBeat reports that in October the Department of Homeland Security disclosed an order directing federal civilian agencies to exclude DeepSeek from procurement and remove its technology from covered systems, including third-party implementations. It found no publicly disclosed U.S. federal restriction naming Meituan or LongCat-2.5 as of November 9. Absence of a restriction today is not a guarantee about tomorrow, and organizations with government customers typically review such policy before adopting a model from any Chinese developer.
How to evaluate it
VentureBeat recommends a structured trial rather than relying on published scores: test screenshot-driven bug fixes, design-to-code tasks and a text-only repository task as a control; hold tools, repository state and time limits constant; repeat runs to measure success rates; score rendered output and code separately; and keep a second model on the same tasks as a fallback. It also advises confirming data retention, access and deletion terms before sending proprietary code or screenshots to the hosted service.
That last recommendation, keeping a second model tested on the same tasks, is the habit that makes open-weight releases useful. Workspaces such as Metir that put several models side by side make it practical to run the same prompt across a new open-weight option and a current provider before changing any production workflow.
Sources:
Image credits
Header and in-body image: Meituan autonomous delivery vehicle (IMG005), via Wikimedia Commons, licensed CC BY-SA 4.0.