LongCat-2.0

Overview

LongCat-2.0 is Meituan's open-weight flagship large language model, released on June 30, 2026. Meituan's model card describes it as "a large-scale MoE language model with 1.6 trillion total parameters and ~48 billion activated per token." The context window is one million tokens, and the weights are MIT-licensed.

The headline is not the model — it is the silicon. Meituan states that "both the full training run and the large-scale deployment are built entirely on AI ASIC superpods," and that "pretraining spans millions of accelerator-days across more than 35 trillion tokens, with no rollbacks or irrecoverable loss spikes — demonstrating that we have the capability to conduct frontier-scale training on alternative hardware platforms." The training section is blunter still: "LongCat-2.0 is pre-trained on over 50K AI ASICs." Meituan's own Chinese share copy for the blog reads 训练全程由国产芯片完成 — the entire training process was completed on domestic chips. South China Morning Post reports Meituan's framing as "the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster," contrasting it with DeepSeek V4-pro, which used home-grown chips for inference only.

Meituan does not name the ASIC vendor. Its only direct reference to the incumbent is a concession: "Compared to the mature Nvidia GPU ecosystem, the supporting software community is still less developed. We have therefore put significant effort into building a stable, secure, and scalable infrastructure." The blog notes that "our accelerators have significantly less per-device memory than an H800 (80 GB)," which drove a 6D parallelism scheme, ZeRO-1, selective recomputation and allocator-level offloading.

A note on parameter counts

Meituan says 1.6T. Hugging Face's API reports 1,775,560,491,136 parameters for meituan-longcat/LongCat-2.0. Both are correct, and the gap is not a rounding error.

The published checkpoint bundles a 3-step Multi-Token Prediction draft module under the model.mtp. prefix. Summing the tensor shapes declared in the headers of the 194 safetensors shards, that module accounts for 136,943,683,712 parameters — of which 134,978,871,296 are a second, replicated copy of the 135B N-gram Embedding table (oe_embed_tokens*). Remove it and 1,638,616,807,424 parameters remain: the ~1.6T Meituan quotes. Meituan publishes no reconciliation of the two figures; the arithmetic above is derived from the shard headers, since the shipped model.safetensors.index.json carries only a byte-size total and a weight-to-file map, not shapes.

Capabilities

• LongCat Sparse Attention (LSA): Meituan's rework of DeepSeek Sparse Attention, targeting "the output discontinuity and quadratic scoring bottleneck of the Lightning Indexer." Three orthogonal pieces — Streaming-aware Indexing, Cross-Layer Indexing, Hierarchical Indexing — each independently switchable.
• Native 1M-token context: Meituan reports training "on hundreds of billions of tokens of 1M-context data," using an all-gather context-parallelism scheme scaled "to over 512" to enable "native 1M-length training."
• N-gram Embedding: 135B parameters spent on an embedding table expanded "by roughly 100× through N-gram token combinations" (n-gram size 5) rather than on more experts — Meituan argues MoE sparsity had "already reached approximately 97%," past the point where extra experts pay.
• Agentic coding: Vendor-reported 70.8 on Terminal-Bench 2.1 and 59.5 on SWE-bench Pro, both run through Claude Code as the harness.
• Multi-Token Prediction: A 3-step MTP draft module for speculative decoding; all three draft steps share a single LSA indexing pass.
• Harness integration: The model card claims deep integration with "Claude Code, OpenClaw, and Hermes"; the API docs add Kilo Code, OpenCode, Codex, ZCode, Cline, Cherry Studio, CodeBuddy and WorkBuddy.
• Dual API surface: Meituan serves both an OpenAI-compatible /v1/chat/completions endpoint and an Anthropic-compatible /v1/messages endpoint.

Technical Specifications

• Hugging Face repository: meituan-longcat/LongCat-2.0 (plus -FP8 and -INT8 variants)
• Total parameters: 1.6T per Meituan; 1,775,560,491,136 in the safetensors index (see the note above)
• Active parameters: ~48B per token (Meituan). Widely circulated "33B–56B dynamic range" figures appear in press coverage, not in Meituan's model card or blog
• Architecture: Mixture-of-Experts transformer, building on LongCat-Flash (arXiv:2509.01322)
• Experts (config.json): 768 routed experts plus 128 identity zero-experts, top-12 routing, 38 layers, hidden size 8192
• Attention: MLA with the LSA indexer (index_topk 2048, cross-layer indexing every 2 layers)
• MTP: 3 draft steps, weights replicated
• Pretraining: >35T tokens, >50K AI ASICs, Muon optimizer
• Context window: 1M tokens; max output 128K tokens (API docs)
• License: MIT

Note that the shipped config.json declares max_position_embeddings: 262144 while its YaRN rope_scaling block (original_max_position_embeddings: 8192, factor: 120) implies roughly 983K. Serving the full advertised million tokens on the open weights is not a matter of loading the config as-is. Meituan does not publish a knowledge cutoff, so this page does not state one.

Use Cases

• Long-horizon coding agents: The model is tuned for and evaluated through agent harnesses, not chat. Terminal-Bench 2.1 and SWE-bench Pro are the benchmarks Meituan leads with.
• Repository-scale code understanding: A 1M-token context plus LSA is aimed squarely at reading a whole codebase in one pass.
• Drop-in substitute for Anthropic-format tooling: Because Meituan exposes /v1/messages, tools written against Claude Code's API shape can be pointed at LongCat with a base-URL swap.
• Cost-sensitive high-volume inference: At $0.30/$1.20 promotional and $0.75/$2.95 list, it undercuts Western frontier pricing by roughly an order of magnitude.
• Sovereign-compute deployments: The most interesting property for some buyers is not the model but the proof that a 1.6T model can be pretrained off the NVIDIA stack.
• Research on sparse attention: LSA and N-gram Embedding are described in enough detail to reimplement, and MIT weights make ablation possible.

Performance / Benchmarks

All numbers below are from Meituan's model card. Meituan's caveat: "Unless noted with *, all scores are measured in-house under a unified harness." Scores marked * for competitors are "cited from the model's official report" — so this table compares Meituan's own runs against other vendors' self-reported numbers, which is not an apples-to-apples comparison.

The harness footnotes live in Meituan's tech blog rather than the model card. The SWE-bench note reads "Evaluated via Claude Code; per sandbox instance 4c8g; inference params temperature=1.0, top_k=-1, top_p=1; problematic tasks corrected"; Terminal-Bench 2.1 used an 8c16g sandbox and a six-hour agent timeout. Of the three "general agent" benchmarks, Meituan describes RWSearch as "an in-house objective benchmark for search agents"; FORTE is published by AGI-Eval, not Meituan. The 59.5 on SWE-bench Pro has not been independently replicated.

Independent placements

There are none. Artificial Analysis's LongCat provider page states plainly: "We are not currently tracking any models from LongCat." On arena.ai, longcat-2.0 exists in the battle registry (served via SiliconFlow) but carries no Elo rating; the only ranked Meituan entries are the older longcat-flash-chat (1401.1) and longcat-flash-chat-2602-exp (1435.59). Note that arena.ai and Artificial Analysis use different scales and are never comparable.

Limitations

• Adoption is the weak spot. Three days after the weights landed, meituan-longcat/LongCat-2.0 shows 385 downloads and 147 likes on Hugging Face. For a 1.6T MIT-licensed flagship, that is a rounding error next to DeepSeek or Qwen releases.
• Not in OpenRouter's live catalogue. As of 2026-07-08 the OpenRouter models API returns 344 models and none of them are Meituan's, under any name.
• Every benchmark is vendor-reported, run in-house, against competitors' self-reported figures. No Artificial Analysis score, no arena.ai Elo, no third-party SWE-bench Pro replication.
• The "Owl Alpha" story is only half-confirmed. Meituan's own launch post on X calls LongCat-2.0 "the full model behind Owl Alpha on @OpenRouter," but the two-month stealth run and the traffic rankings are media reporting. OpenRouter has said nothing, and the listing no longer exists to inspect.
• Circulating usage numbers are unusable. The most-quoted token-volume figure — 10.1 trillion tokens a month — cites no source and contradicts the per-day average given in the same sentence of the same article.
• The ASIC claim is unaudited. Meituan names no chip vendor, publishes no cluster telemetry, and no independent party has verified that no NVIDIA hardware touched the run. It is a strong claim resting entirely on the vendor's word.
• Self-hosting is brutal. 1.78T parameters in BF16 is ~3.55 TB of weights across 194 shards. FP8 and INT8 checkpoints exist, but this is a datacentre-scale model, not a workstation one.
• Open weights, closed everything else. MIT covers the weights. Training data, data mixture, and the training stack are not published.

Pricing & Access

Self-hosting is free. MIT-licensed weights at meituan-longcat/LongCat-2.0, mirrored in full on ModelScope (all 194 shards). The GitHub repository carries only the model card and licence, not the weights. Commercial use, modification and redistribution are permitted with no field-of-use or geographic restriction.

Hosted access is sold through the LongCat API Platform. Meituan's pricing page publishes the rate card in US dollars (the Chinese-language docs carry a parallel yuan card: ¥5 / ¥0.10 / ¥20, discounted to ¥2 / ¥0.04 / ¥8):

The discounted column is explicitly labelled "limited-time" by Meituan — the Chinese page calls it a launch promotion — and the page adds: "Prices are subject to change. The prices shown on the LongCat API Platform and your billing records are authoritative." On Pay-As-You-Go, cached input is not free, despite several outlets reporting that it is; on the prepaid Token Pack it effectively is, because Meituan's Token Pack FAQ says "Cache hits are free and do not count toward package token usage."

Billing comes in two forms per the change log: a Token Pack (prepaid quota, valid 30 calendar days) and API Pay-As-You-Go. Endpoints are https://api.longcat.chat/openai (OpenAI format) and https://api.longcat.chat/anthropic (Anthropic format), with a 1M-token context and 128K max output. Rate limiting returns HTTP 429.

Serving: GPU deployment via the SGLang cookbook; NPU deployment via Meituan's own SGLang-FluentLLM fork.

Consumer access is at longcat.ai, Meituan's chat product.

Ecosystem & Tools

• meituan-longcat/LongCat-2.0 on Hugging Face — MIT weights, model card, benchmark table
• LongCat-2.0 tech blog — LSA, N-gram Embedding, ASIC superpod infrastructure, MOPD post-training
• LongCat API Docs — quick start, endpoints, change log, tool guides
• SGLang cookbook entry — supported GPU serving path
• SGLang-FluentLLM — Meituan's NPU serving fork
• ModelScope mirror — weights for users inside China
• Agent harnesses — Claude Code, OpenClaw, Hermes Agent, OpenCode, Kilo Code, Codex, Cline, Cherry Studio, CodeBuddy, WorkBuddy, all documented on the platform docs site

Community & Resources

• China claims biggest AI model trained on local chips, as Meituan releases LongCat-2.0 — South China Morning Post, June 30, 2026
• China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips — SiliconANGLE
• Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia — The Decoder
• Meituan LongCat's launch post on X — June 30, 2026; Meituan's own confirmation that LongCat-2.0 is "the full model behind Owl Alpha on @OpenRouter"
• LongCat-2.0: The Stealth AI Model That Was Quietly Topping OpenRouter All Along — Decrypt, July 1, 2026; source of the two-month stealth-run and traffic-ranking details, which Meituan does not corroborate
• Artificial Analysis: LongCat — currently tracking no LongCat models
• Compare with DeepSeek V4, Kimi K2.6, GLM-5.2, Ling-2.6-1T, Tencent Hy3, and MiniMax-M3