Chinese food-delivery and local-services company Meituan has released LongCat-2.5, a new multimodal model with 1.6 trillion total parameters, 49 billion of which are active, and says it is now available through its API.
Model release · Meituan
Meituan launches LongCat-2.5, a 1.6-trillion-parameter multimodal model open to developers via API
The model uses a sparse design: only about 49 billion of its 1.6 trillion parameters run for any given input. Meituan positions it against leading proprietary models, though independent benchmarks are not yet available.
49B
Active parameters per computation
~3%
Share of parameters active at once
Total vs. active parameters, drawn to scale
Each block represents about 50B parameters. The total model is roughly 32 blocks; the active share is 1 block.
49B
Active per computation
Takeaway: the total model is about 33 times larger than the portion that runs for any single input.
What the company claims
Frontier-tier positioning
Compared with GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8 and Claude-Opus-5
Available to developers through its API
What remains unverified
Independent benchmark results
Third-party evaluations
Full architecture, training data, context length and pricing
Reading the numbers
A sparse design aims to keep running costs closer to a much smaller model while keeping large-model capacity.
Performance claims against rival models cannot yet be verified.
Cost, measured quality and real-world uptake will decide whether the model gains traction.
Developers and researchers are expected to test it closely in the coming weeks.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...