On November 1, Meituan's LongCat team released LongCat-2.5 with a full technical report; the model went live on the LongCat API platform the next day (official release page). The spec sheet barely moved: 1.6T total parameters, roughly 48B active per token, 1M context, 128K output — the same envelope as the September preview. What changed is the training target: 2.0 was built to write code, 2.5 is built to finish projects — across a codebase, a terminal, a browser, a spreadsheet, or a design canvas. And it can see.
Same parameters, different scores
The official comparison against LongCat-2.0 tells the story: Terminal Bench v2.1 70.8 to 88.4, v4.0 3.0 to 38.5, DeepSWE v1.1 9.7 to 64.6, PaperBench 45.2 to 82.1, BrowseComp 79.9 to 91.9. Same skeleton, multi-step scores multiplied several times over — and the team is explicit in the report that long-horizon agentic RL, not pretraining scale, did the work. To its credit, the report does not hide the gaps: in the comparison set Claude Opus 5 tops Terminal Bench v4.0 at 53.9, about 15 points ahead; DeepSeek V4.1-Flash leads DeepSWE v1.1 and Terminal Bench v2.1; pure-vision results sit mid-pack (VideoMME 85.2 against Kimi K3's 90.0). The official summary is blunt: competitive breadth with a visual-coding edge, not yet dominant on the hardest engineering tasks.
Vision in the backbone
"Native multimodal" here means no bolted-on vision encoder: images become multi-level discrete tokens — RVQ over native-resolution SAE features, the LongCat-Next lineage — so text and vision share one token stream and one objective. Visual tokens are precomputed and cached offline, which means the vision encoder never enters the training graph at all. On a roughly 3B-active probe model, t-SNE shows visual and text embeddings interwoven rather than clustered apart. The shipped form is "visual coding": read a screenshot, chart, UI, or 3D viewport inside the same reasoning loop that writes the code.
W4A4 on domestic accelerators
On the infrastructure side, long-context training and W4A4 inference are both tuned for domestic Chinese accelerators. Inference quantizes expert weights and activations to 4 bits, with FlatQuant-style learnable equivalent transformations recovering accuracy; the company says the quantized model matches the unquantized one on every evaluated downstream task. Against W8A8 in online stress tests: prefill throughput +43.3%, decode +42.77%, TTFT down 15.91%. On the training side, the FSDP configuration delivers 86% higher per-card throughput than a DDP baseline on long-sequence SFT.
Open weights stop at 2.0
The most telling detail is distribution: no open-weight release has been announced. That breaks the family default — Flash-Chat, Flash-Thinking, LongCat-2.0 (MIT), Next, and Video-Avatar all shipped weights. For self-hosters, the practical reality is that 2.0 remains the newest usable open weight, and 2.5 has to be treated as a hosted service with the usual latency, data-residency, and pricing dependencies. Third-party coverage (TechNode via AI Weekly) confirms the November 2 availability and the domestic-chip training angle, but whether weights follow is a question for the next releases.
So: when parameters stop growing and Terminal Bench v4.0 still jumps from 3.0 to 38.5, the "more pretraining tokens" story yields to whoever can actually run RL in real environments. Wait for weights, or swap a base URL and try the API?