AI News
All stories
Reported

LongCat 2.5 open model reaches 1.6 trillion parameters with 48B active

Matthew Berman says LongCat 2.5 is a 1.6-trillion-parameter multimodal model with 48 billion active parameters that can be downloaded.

Open Source · · 2 min read

Meituan's LongCat team has released LongCat-2.5, an open-weights multimodal model with 1.6 trillion total parameters and about 48 billion activated per token. The weights are available under the MIT License. Matthew Berman, who covered the release in a video, called it a 1.6-trillion-parameter model with 48 billion active parameters that can be downloaded. He said he had not heard of the company before.

The active count and the total count measure different things. The model card describes LongCat-2.5 as a mixture-of-experts model, in which a router sends each token to a small subset of the network's experts. Per-token compute therefore scales with the 48 billion activated parameters. All 1.6 trillion parameters must still be stored somewhere to serve requests. Berman said that at 1.6 trillion parameters most people will not be able to run the model locally. The model card's own metadata lists the safetensors size as 1.8T params, not 1.6T, and the page does not explain the difference.

The card describes the model as built for long-horizon agentic tasks, combining visual understanding with agentic coding. It covers complex codebase development, full-project builds, creative design and 3D tasks. Berman's claims of stronger visual understanding and strength in creative design and 3D can be checked against the card's benchmark table, which is mixed. On 3DCodeBench (Text→3D), LongCat-2.5 scores 55.4, above the 54.3 listed for Claude-Opus-5 and the 53.4 listed for GLM-5.3. On VideoMME, however, it scores 85.2 against 89.5 cited for GPT-5.6-Sol and 90.0 cited for kimi-k3. The card marks scores with an asterisk when they are cited from public reports. It says unmarked scores were measured in-house and are indicative only. The 3D result is unmarked, so it is the company's own measurement, and the margin over the next model is about one point.

The card also reports quantization work for serving. LongCat-2.5 applies W4A4 quantization to its expert weights and activations, which the card says enables native INT4 computation on domestic AI accelerators. Compared with a W8A8 MoE baseline, the card reports 43.3% higher prefill throughput and 42.77% higher decode throughput in online stress tests. Those figures come from the company's own tests. The card does not name the hardware or the workload beyond that description.

What changed

Meituan's LongCat team released LongCat-2.5, an open-weights multimodal model with 1.6 trillion total parameters and about 48 billion activated per token.

Type: New capability · Applies to: LongCat-2.5

Why it matters

Very large open models widen the choice for organisations with heavy inference infrastructure, though most teams will need hosted access.

Original source

Where we saw it

1 source

Forward Future (Matthew Berman)

LongCat

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works