LongCat 2.5 Launches: 1.6T Multimodal Agent Model

On November 1, 2026, Meituan's LongCat team published "Introducing LongCat-2.5" on its official tech blog; the next day, November 2, LongCat 2.5 went GA on the API platform, with video input and Codex access arriving at GA. We use November 2 as the release date — that is when the model became generally available, and eesel, toolpilotpro and the China AI Index all use November 2 as the release date too. This is the second-generation flagship of Meituan's long-horizon agent line: the same 1.6T/48B base as LongCat 2.0 (GA June 30), now with native vision and large-scale long-horizon reinforcement learning, promo pricing identical to DeepSeek V4.1 Flash's peak rate, and weights released under MIT on November 4.

Relation to 2.0: same base, plus vision and agent training

LongCat 2.0 went GA on June 30, 2026 (its Hugging Face repo was created July 5) as a text-only model with 1.6T total parameters and about 48B active. LongCat-2.5-Preview arrived on the API platform on September 25 with image-only input, no published benchmarks and no parameter disclosure, and was free inside OpenCode for two weeks. The full 2.5 does not swap the base — 38 layers, hidden size 8192, 768 routed experts, topk 12, MLA attention, a 163,840-token vocabulary — it adds two things on top of it: native vision and long-horizon agent training.

On vision, 2.5 takes text, image and video input and produces text-only output. Images support PNG/JPG/JPEG/WebP/GIF, up to 10MB each and 50 per request; video is up to 100MB and under 3600 seconds per clip, 10 clips per request, public URLs only; SVG and HEIC are unsupported. The vision mechanism uses multi-level discrete tokens from LongCat-Next, with no separate vision encoder: native-resolution tokenization, with visual embeddings learned jointly with the backbone. Context is 1M tokens (API documentation figure; the config's base is 512K, extended via YaRN). Maximum output is 128K — though another documentation page allows max_tokens up to 262,144, and the two figures disagree, pending official clarification.

Training and inference: long-horizon RL, MOPD distillation, domestic accelerators

The training headline is long-horizon reinforcement learning: tasks filtered from GitHub PRs with F2P/P2P validation, real failure trajectories converted into reproducible tasks, anti-reward-hacking measures, and penalties on repeated reasoning and repeated tool calls. Distillation uses the MOPD multi-teacher scheme: it keeps 2.0's agent, reasoning and interaction expert groups, with full-vocabulary KL and Tree Attention. Training ran on domestic accelerators (under 80GB per card), with supernode-aware parallelism and dual-shard FSDP — long-sequence SFT throughput per card is 86% above the DDP baseline.

The inference story is W4A4 quantization: INT4 expert weights plus INT4 activations. Against a W8A8 INT8 baseline, the official blog reports online load-test gains of +43.3% prefill throughput, +42.77% decode throughput, −15.91% time to first token and −19.72% time per output token; the GMM kernel averages 1.73×, and W4A4KV8 decode throughput is +92% versus W8A8KV16. All of these are Meituan's own measurements; independent reproductions have not appeared yet.

Pricing and weights: promo pricing matching DeepSeek, MIT weights out

Promo pricing is $0.30 input, $0.006 cached input and $1.20 output per 1M tokens (¥2 / ¥0.04 / ¥8), all marked "Discounted Price (limited-time)" — with no end date and no published list price. The $0.30/$1.20 rate is exactly DeepSeek V4.1 Flash's peak price; the benchmark head-to-head is in the benchmarks article.

The weights were released on Hugging Face on November 4 under MIT: the full checkpoint is 3.3TB across 210 shards, the W4A4 checkpoint 1.24TB across 208 shards, with a ModelScope mirror. Reports from November 1–3 (and even toolpilotpro's November 4 updated comparison) called the weights "coming soon" and the model API-only — that claim has been out of date since November 4, and we follow Hugging Face.

The official blog includes eight end-to-end generated demos: on software development, a real-time fluid simulation, an interactive 3D artifact exhibition, a city-scale model and miniature scene assets; on professional office work, a business operations analysis, a customer feedback analysis report, a Chinese-colour aesthetics page and an event invitation generator. All are official demos, not independently reproduced, and the demo videos were not yet live in the first days after launch.

What hasn't landed yet

As of November 13, no independent body has scored 2.5: Artificial Analysis has not listed it, and ArtificialWatch published a "not rated yet" report the same day. The data policy is also blank: the FAQ and billing pages do not state prompt retention periods or whether API traffic is used for training — something customer-service and compliance users should note. The discount has no end date, and Meituan has not said what pricing follows it. Separately, 2.0 and 2.5-Preview have disappeared from the official Supported Models table, but with no dated sunset announcement — we write only that they are "no longer on the official model list," not that they are discontinued.