AI · 17 November 2026
Meituan's LongCat-2.5 agent model: 1.6T parameters on Huawei Ascend
Meituan has released model weights for LongCat-2.5, a 1.6-trillion-parameter agent model that scores 88.4 on Terminal Bench, just behind Claude Opus 5, according to reporting by The Decoder. The sample deployment runs on 194 Huawei Ascend chips.
What happened
Chinese technology group Meituan has released the model weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and, according to its reporting, trained entirely on Chinese hardware. The release was covered by The Decoder, which framed it around the model's hardware sourcing.
On Terminal Bench, a benchmark used to assess agent performance, LongCat-2.5 scores 88.4 points. That places it just behind Claude Opus 5, the model developed by Anthropic, according to the same report.
The sample deployment described in the coverage runs on 194 Huawei Ascend chips. Meituan's material, as reported, makes no reference to Nvidia GPUs, which is the focus of the headline the story has carried across outlets.
Why it matters
The significance here lies less in a single benchmark score than in what the release signals about agent-focused AI. A model of this scale aimed squarely at agentic tasks, meaning systems that plan and carry out multi-step work rather than simply answering prompts, changes what organisations can build on. Leaders evaluating AI platforms will need to weigh capability alongside where and how a model runs.
For digital transformation teams, the deployment detail matters as much as the benchmark. A model demonstrated on a specific accelerator stack raises practical questions about infrastructure choices, procurement, portability and the cost of running agents at scale. Operators in any market will want to understand which parts of their AI roadmap depend on particular hardware, and how quickly that dependency can change.
By the numbers
- 1.6 trillion parameters in the LongCat-2.5 model.
- 88.4 points on Terminal Bench, just behind Claude Opus 5.
- 194 Huawei Ascend chips used in the sample deployment.
The Renascence take
Most of the coverage is fixated on the hardware story, but the experience question is more useful. An agent is only as good as the service it delivers in the moments that matter, and benchmark rankings say little about whether a customer's problem actually gets resolved. The behavioural principle underneath is simple: people judge a system by its failures and recoveries, not its leaderboard position.
Benchmarks reward capability, but customers reward reliability, clarity and a graceful handover when the agent gets stuck. The real test of an agent model is not whether it lands near the top of a table, but whether it keeps a customer's trust through a messy, multi-step request. A customer-obsessed operator should therefore run pilots on their own journeys, measure task completion and escalation quality, and treat hardware choices as a service-continuity decision rather than a headline.
Sources
This briefing was written by our Newsdesk, synthesising reporting from the outlets below. Follow the links for the original coverage.
FAQ
Questions we get on this topic
More in AI
Stay ahead of CX
Get the signal, not the noise.
The stories shaping customer experience — plus the Journal and Experience Loom — in your inbox.


