Meituan's LongCat team released LongCat-2.5 on November 1, adding a full technical report and evaluation table to the preview that went live on September 25. The model can already be tried on the official website, and API access is connected, but there is no open-source release yet: the GitHub and HuggingFace links on the page both say "coming soon". That differs from LongCat-2.0 at the end of June, which was open-sourced at launch.
The model's size is unchanged, at 1.6 trillion total parameters with about 48 billion activated per token. What has changed is the direction of its capabilities: native multimodality on one side, and agentic coding for long-horizon tasks on the other.
Big gains over 2.0, a mixed picture against the leaders
The official table lists more than 30 benchmarks. We picked a few representative ones:
| Benchmark | LongCat-2.5 | LongCat-2.0 | GPT-5.6 Sol | Opus 5 | GLM-5.3 | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|---|---|---|---|
| Terminal Bench v4.0 | 38.5 | 3.0 | 37.3 | 53.9 | 41.8 | 12.6 | 31.2 |
| Program Bench | 25.5 | 2.5 | 23.0 | 37.0 | 19.0 | 17.5 | 20.3 |
| DeepSWE v1.1 | 64.6 | 9.7 | 72.7 | 68.8 | 66.9 | 67.5 | 74.2 |
| NL2Repo | 60.5 | 40.9 | 56.8 | 75.3 | 58.0 | 58.0 | 65.4 |
| BrowseComp | 91.9 | 79.9 | 90.4 | 90.8 | 81.0 | 91.2 | 91.5 |
| 3DCodeBench | 55.4 | - | 52.3 | 54.3 | 53.4 | 49.3 | 53.3 |
A few signals stand out. On projects that require running a complete engineering task over a long stretch, the improvement over 2.0 is largest: Terminal Bench v4.0 and Program Bench both rose more than tenfold. On BrowseComp, ResearchRubrics, ALE and 3DCodeBench, it posts the highest score in the official table. On most coding benchmarks it still trails Opus 5, and its 64.6 on DeepSWE is the lowest among the competitors in the table above.
Three caveats apply when reading this table. First, the comparison set does not include the GPT-6 series, Opus 5.5 or Fable 5.1, so the model has not been measured against these newer systems. Second, competitor scores marked with an asterisk are taken from public reports, while unmarked ones were measured by Meituan itself, and the company labels the figures "for reference only". Third, the harnesses and parameters differ from one benchmark to another. No independent replication has been seen so far, so for now these scores can only be treated as Meituan's own account.
How vision gets into the backbone
For multimodality, images are encoded into multi-level discrete tokens, so that the backbone model handles vision and text in the same token format. The visual tokenizer is fixed, which means token IDs can be computed in advance and cached, and the training graph no longer contains a vision encoder. The team says this simplifies long-sequence packing and also improves load balancing in distributed training.
Officially, this capability is presented together with coding under the name Visual Coding. The demos include a real-time fluid simulation, a 3D display of cultural relics and a city sand-table model.
Long-horizon ability is built up through reinforcement learning. Training tasks are synthesized from public GitHub repositories and internal code scenarios, and trajectories from real agent runs that failed are also turned into reproducible training problems. To keep the model from gaming the tasks, the environment cuts off network access and clears git history and file caches, and an LLM judge watches the trajectories, applying extra penalties for repeated thinking and repeated tool calls.
The compute-side numbers
The report spends a good deal of space on domestic accelerator cards, repeatedly citing per-card memory of under 80GB as a constraint. The two published optimization results are:
- Long-sequence training: with an FSDP configuration designed for supernodes, per-card throughput is 86% higher than the DDP baseline.
- W4A4 inference: with both expert weights and activations quantized to 4 bits, live load tests against a W8A8 baseline show prefill throughput up 43.3% and decode up 42.77%, while TTFT and TPOT fall by 15.91% and 19.72% respectively.
The report does not say which vendor's cards were used, how many, or what the training cost. The technical blog also lists no API price for the official release; the preview was offered at a limited-time price, and any new pricing should be checked against announcements on the open platform. No open-source date has been given either, and until the weights are released, outsiders cannot verify this evaluation table for themselves.
Sources: LongCat official technical report and evaluation table, CocoLoop, ITHome. All benchmark scores follow the figures published by the company; competitor scores are partly taken from public reports and partly measured by Meituan.