Meituan releases LongCat-2.5 multimodal agent API with 1.6T parameters

• LongCat-2.5 is a native multimodal model with 1.6T total parameters and roughly 48B active per token, according to the launch announcement.
• API access combines a 1M-token context with promotional pricing of $0.30 per million uncached input tokens and $1.20 per million output tokens, as WesRoth reported.
• Training combines long-horizon agentic RL with specialist teachers; Meituan's technical overview highlights shared prefill, KV-cache reuse and Tree Attention.
• Cache billing differs across launch materials: the API announcement promises free cache hits, while the pricing page lists $0.006 per million cached input tokens.

Meituan's technical write-up includes a visualization of interwoven image/text embeddings, but that experiment used a smaller model with about 3B active parameters. The Claude Code configuration maps Opus, Sonnet and small/fast model slots to LongCat-2.5, while the Codex example disables web search.

The promotional API pricing is aggressive for a 1M-context model. Meituan lists three limited-time rates on its pricing page:

| Token type | Price per million tokens |
| --- | ---: |
| Uncached input | $0.30 |
| Cached input | $0.006 |
| Output | $1.20 |

The cached-input charge conflicts with the launch offer's promise that cache hits remain free. The documentation says platform prices and billing records are authoritative.

Meituan documents connections to three existing agent harnesses:

• Claude Code: ~/.claude/settings.json sets ANTHROPIC_BASE_URL to https://api.longcat.chat/anthropic and ANTHROPIC_MODEL to LongCat-2.5, per the integration guide.
• Codex: ~/.codex/config.toml uses https://api.longcat.chat/openai/v1, model LongCat-2.5 and wire_api = "responses", according to the Codex guide.
• Hermes Agent: a custom provider uses the same OpenAI-compatible base URL and model slug, as shown in the Hermes guide.

LongCat-2.5 builds on LongCat-2.0 with a unified token-based path for images and text. Meituan describes three components in its visual-modeling explanation:

1. Native-resolution tokenization: a Semantic-and-Aligned Encoder produces image features; residual vector quantization converts them into multi-level discrete tokens.
2. Offline preprocessing: a fixed tokenizer lets training pipelines precompute and cache visual token IDs, removing the vision encoder from the backbone's training graph.
3. Joint embeddings: visual codebook embeddings are learned alongside the backbone.

Known image-token counts also let the training system balance modality proportions across data-parallel ranks before execution.

Training tasks come from filtered GitHub pull requests, linked issues and internal engineering work, according to Meituan's training write-up. Agents annotate tasks and synthesize interactive environments.

• Verification: fail-to-pass tests check the requested fix; pass-to-pass tests check existing behavior. Live rollout pass rates guide re-labeling and filtering, while false-positive and false-negative analyses drive repairs and extensions to the fail-to-pass tests.
• Asynchronous rollouts: once a training batch fills, unfinished trajectories are interrupted, saved and resumed during the next sampling phase. Tokens can therefore come from different historical checkpoints; log-probability clipping and per-token gradient rescaling limit the resulting staleness bias.
• Exploit controls: task environments have network access disabled, git history removed and cached files cleared. An LLM judge audits trajectories for exploits.
• Output controls: additional penalties target repetitive reasoning and repeated tool calls. Reward redistribution among positive samples targets overthinking and excessive response-length growth.

The MOPD stage retains LongCat-2.0's expert-group setup for agent capabilities, reasoning and interaction. Students sample online, and specialist teachers supply training signals through full-vocabulary KL divergence, according to the distillation description.

Three mechanics control distribution shift and duplicated computation:

1. Per-cut prefixes: teacher responses are divided into fixed-length segments, with student self-sampling capped at the same length.
2. Shared inference: prefixes from one teacher trajectory run on the same inference engine, sharing prefill computation and KV cache.
3. Tree Attention: the teacher trajectory becomes the trunk and student samples become branches, allowing all grouped samples to train in one forward and backward pass.

Meituan quantizes expert weights and activations to four bits and executes their matrix multiplications on native INT4 units. In its domestic-accelerator stress tests, prefill throughput rose 43.3% and decode throughput rose 42.77% against a W8A8 INT8 baseline.

Meituan also describes three optimizations for long-context training rollouts:

• Effective cache capacity: prefill partitions queries by token and distributes KV by page, avoiding KV duplication associated with tensor parallelism in MLA. HiCache releases unreferenced host copies after KV returns to device memory.
• Cross-layer prefetching: odd-numbered layers reuse the preceding even layer's selected indices. Their full KV can reside in host memory, with selected entries prefetched during even-layer computation; Meituan reports nearly doubling decode KV capacity without increasing time per output token.
• Capacity-aware scheduling: the router admits requests according to combined L1/L2 cache capacity and token counts, prioritizing groups closer to completion to release cache sooner and avoid repeated prefix eviction.

Meituan reports substantial gains over LongCat-2.0 in its evaluation table:

| Benchmark | LongCat-2.0 | LongCat-2.5 | Score change |
| --- | ---: | ---: | ---: |
| Terminal Bench v4.0 | 3.0 | 38.5 | +35.5 |
| DeepSWE v1.1 | 9.7 | 64.6 | +54.9 |
| BrowseComp | 79.9 | 91.9 | +12.0 |

The setup used Claude Code v2.1.280 with a 512K sequence limit for Terminal Bench, versus mini-swe-agent with a 1M limit for DeepSWE. BrowseComp used a custom harness with Verify and Discard-All modules and up to 1,024 exploration rounds.

Competitor scores mix public reports and Meituan's own runs. On DeepSWE, LongCat-2.5's 64.6 trails the table's 72.7 for GPT-5.6-Sol and 68.8 for Claude-Opus-5.

Early website feedback was less enthusiastic: teortaxesTex reported poor Russian-language results and disputed the frontier-model comparisons.

Meituan says its showcases were generated end to end by LongCat-2.5, including these interactive visual-coding projects:

Two office-work demos extend the launch beyond coding:

• Business analysis: multi-channel data integration and metric reconciliation feed profit and ROI analysis, followed by marketing, supply, inventory and budget recommendations in the business-analysis demo.
• Customer feedback: duplicate and follow-up detection feed severity and urgency prioritization, with an interactive HTML report linking back to original messages in the feedback demo.

Meituan promised 10 million tokens for new signups and an additional 10 million for existing users, credited automatically.

The credit did not arrive for teortaxesTex, who reported receiving nothing after signing up. Meituan apologized, said the team was investigating and offered to follow up by direct message.

The reply refers to “1,000M” new-user tokens, a different amount from the 10 million advertised in the launch announcement.