Meituan

Executive Summary

What it is: Meituan's LongCat-2.0 (1.6T/48B MoE, 1M context, MIT-licensed weights, trained on Chinese AI ASICs) remains current, served via a pay-per-token API at longcat.chat ($0.30/$1.20 per MTok discounted, China-centric payments) with harness integrations including Claude Code and OpenCode. Source: https://github.com/meituan-longcat/LongCat-2.0

What to watch out for: Two months of total silence since the June 30 launch: no roadmap, no API updates, no community engagement, and international payment/access barriers unchanged. Source: https://github.com/meituan-longcat/LongCat-2.0

Bottom line: LongCat-2.0 is a credible self-hosted open-weight option (GPU and NPU deployment documented) and proof that trillion-parameter training works on domestic Chinese chips, but as a supplier Meituan shows no ongoing operational signal; keep tracking light. Source: https://github.com/meituan-longcat/LongCat-2.0

Key Terms

  • LongCat-2.0 - Meituan's open-weight flagship: 1.6T total / ~48B active MoE, trained entirely on AI ASIC superpods across 35T+ tokens, 1M context, MIT-licensed weights. Source: GitHub – Longcat 2.0
  • LongCat Sparse Attention (LSA) - Meituan's sparse-attention variant (Streaming-aware, Cross-Layer, and Hierarchical Indexing) addressing DSA's output discontinuity and quadratic scoring. Source: GitHub – Longcat 2.0
  • N-gram Embedding - 135B parameters of sparse-dimension expansion orthogonal to MoE, inherited from LongCat-Flash-Lite. Source: GitHub – Longcat 2.0

Latest Changes

Changes since the 2026-07 report.

July watch items verified:

  • LongCat-2.1 or successor: no signal. No new model, API change, or blog post appeared in August; the repo's benchmark figures remain dated June 29 and the June 30 release stands as current. Source: GitHub – Longcat 2.0

August status: no updates. No pricing, model, or product changes at longcat.chat or in the GitHub/HuggingFace channels. The historical note stands: LongCat-2.0 was previously surfaced as the OpenRouter stealth model "Owl Alpha" before official release, part of the Chinese-lab stealth-preview pattern that recurred with Zhipu's suspected "OX Alpha" this month. Source: GitHub – Longcat 2.0

Plans

No subscription plans. Pay-per-token API at longcat.chat (discounted $0.30/$1.20 per MTok uncached input/output, carried over; payment via Alipay/WeChat only). Source: Longcat – Platform

API Pricing

Carried over unchanged: $0.30/$1.20 per MTok (discounted, uncached). Source: Longcat – Platform

Model Performance / Benchmarks

Standing June figures (in-house harness unless starred): Terminal-Bench 2.1 70.8, SWE-bench Pro 59.5, SWE-bench Multilingual 77.3, GPQA-diamond 88.9, BrowseComp 79.9. Source: GitHub – Longcat 2.0

Latest News

None in August. Source: GitHub – Longcat 2.0

Community Signals

No new significant community threads found in August. The open-weight download ecosystem (HuggingFace/ModelScope GGUF quants) remains the main usage channel. Source: Hugging Face – Meituan Longcat

Enterprise Readiness

Feature Available? Details
SSO (SAML/OIDC) No Not available.
SCIM No Not available.
Audit logs No Not available.
IP indemnity No Not available.
Data residency No China-hosted API only; self-hosting via open weights is the alternative.
HIPAA No Not applicable.
Air-gapped / on-prem Yes MIT weights; GPU (SGLang) and NPU (SGLang-FluentLLM) deployment documented. Source: GitHub – Longcat 2.0
SLA No Not published.
Admin controls (RBAC) No Not available.

Transparency Gaps

Gap Details Severity
No update cadence Two months without any follow-up communication since the June 30 launch; roadmap unknown. Medium
Self-hosting cost of ownership 1.6T weights require multi-node serving; no reference deployment costs published. Low
API internationalization Payment and access remain China-centric. Medium