← Back to feed

Meituan released LongCat-2.5 on November 2, a native multimodal successor to the June LongCat-2.0. It keeps the same shape, 1.6 trillion total parameters with about 48 billion active per token, and adds visual understanding built into the backbone plus a heavier agentic reinforcement learning stage. GitHub and Hugging Face links on the launch page read “coming soon”, so the weights are not yet downloadable.

The benchmark split

All LongCat-2.5 figures are Meituan’s own, run in the Claude Code harness for the coding suites. Competitor numbers are mostly taken from public reports.

• Terminal-Bench 2.1: 88.4, against Claude Opus 5 at 89.1, GPT-5.6-Sol at 88.8 and GLM-5.3 at 88.2. LongCat-2.0 scored 70.8.
• Terminal-Bench 4.0: 38.5, against Opus 5 at 53.9, GLM-5.3 at 41.8 and GPT-5.6-Sol at 37.3. LongCat-2.0 scored 3.0.
• Program Bench (almost resolved): 25.5, against Opus 5 at 37.0 and GPT-5.6-Sol at 23.0.
• DeepSWE v1.1: 64.6, against GPT-5.6-Sol at 72.7 and Opus 5 at 68.8.
• BrowseComp: 91.9, the highest in the table, above Opus 5 (90.8) and GPT-5.6-Sol (90.4).
• HLE with tools: 63.1, against Opus 5 at 64.7.
• MCP Atlas: 83.1, against Opus 5 at 85.8.

The pattern is the story. On the saturating benchmark, a Chinese model is inside a point of the best closed model. On the harder successor, the gap is 15 points. The jump from 3.0 to 38.5 on Terminal-Bench 4.0 over four months is large, but the headroom the new benchmark exposes is still mostly with Anthropic.

Inference on domestic hardware

Meituan says LongCat-2.5 is trained and served on domestic AI accelerators with under 80GB of memory each. For serving, it quantises both expert weights and activations to 4 bits (W4A4) and runs them on native INT4 units. Against a W8A8 INT8 baseline, it reports prefill throughput up 43.3%, decode throughput up 42.77%, time to first token down 15.91% and per-token latency down 19.72%. The company says quality is comparable to the unquantised model on all evaluated downstream tasks.

Caveats

Meituan notes that competitor scores not marked as sourced from public reports are self-run and “indicative only”. Several cells in the table are blank where no comparable number exists. Independent runs will matter, especially on Terminal-Bench 4.0, where the harness version changes results.