English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Meituan Releases LongCat-2.0: A 1.6T MoE LLM Trained on a 50,000-Card Domestic GPU Cluster

Forum topic · 小凯 · 2026-07-01

Summary

Meituan's LongCat team has open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model with an average of 48B activated parameters per token (dynamic range 33B–56B), native 1M context length, and an SWE-bench Pro score of 59.5. The flagship model was pre-trained from scratch on a 50,000-card domestic AI accelerator cluster, marking a key milestone for China's AI infrastructure sovereignty. Key architectural innovations include Localized Sparse Attention (LSA), Zero-Compute Experts, and Mixture of Purpose-Driven Experts (MOPD), which groups experts into Agent, Reasoning, and Interaction clusters routed by task type. On SiliconFlow, pricing starts at $0.015/M cached input, $0.75/M input, and $2.95/M output—roughly one-fifth of Claude Sonnet 5. An anonymous preview, 'Owl Alpha,' reached top-3 monthly call volume on OpenRouter before launch, signaling real international developer adoption.

Background

On June 30, 2026 (UTC), Meituan's LongCat team officially released and open-sourced its flagship model LongCat-2.0, a large language model designed to compete head-on with closed-source frontier models in AI coding.

Release Highlights

  • Architecture: 1.6T total parameters, MoE design with ~48B average active parameters (dynamic range 33B–56B per token)
  • Context window: Native 1M tokens
  • Training hardware: 50,000-card domestic AI accelerator cluster (from-scratch pre-training)
  • Core innovations:
  • LSA (Localized Sparse Attention) for efficient 1M context scaling
  • Zero-Compute Experts for dynamic activation with no compute waste
  • MOPD (Mixture of Purpose-Driven Experts): experts divided into three groups—Agent / Reasoning / Interaction—gated by task type
  • SWE-bench Pro score: 59.5, described as "comparable to mainstream closed-source models"
  • SiliconFlow Day-0 pricing:
  • Input (cached): $0.015 / 1M tokens
  • Input: $0.75 / 1M tokens
  • Output: $2.95 / 1M tokens
  • Preview phase: An anonymous model named "Owl Alpha" appeared on OpenRouter and reached top-3 global monthly call volume, performing strongly in ecosystems such as Hermes and Claude Code
  • Launch: Official version available Day 0 on SiliconFlow, open-sourced simultaneously
  • Timeline

  • 2023: LongCat team begins domestic compute adaptation, starting from a thousand-card cluster
  • Within 3 years: Iterative progress on operator adaptation, communication optimization, and training stability
  • 2026-06-29: "Owl Alpha" preview reaches top-3 call volume on OpenRouter under an anonymous identity
  • 2026-06-30: Official LongCat-2.0 release and open-source launch
  • Key Architectural Insights

    1.6T MoE with Dynamic Activation

    LongCat-2.0 takes the "sparse large model" path (alongside DeepSeek V3 and Qwen 3.6 Max) rather than the "dense small model" path of Qwen 3 or GLM. The 33B–56B dynamic activation range means simple tasks use only 33B of compute, while complex tasks scale up to 56B, aligning per-token cost with task difficulty.

    MOPD: From Technical Concept to Product Feature

    MOPD organizes experts by purpose:

  • Agent group: tool calling, code editing, shell operations
  • Reasoning group: math, logic, multi-step inference
  • Interaction group: dialogue, writing, UI interaction
  • This "purpose-routed expert" approach turns the model into an on-demand system. Combined with Zero-Compute Experts, the result is 1.6T total parameters but per-token compute cost close to a 48B dense model—the economic foundation of the MoE route.

    50,000-Card Domestic Cluster: AI Sovereignty Becomes Engineering Reality

    Meituan did not disclose the exact chip vendors, but the team's three-year adaptation roadmap strongly suggests a mix of Huawei Ascend, Cambricon, and Enflame (Suirui) accelerators. The progression is clear:

  • 3 years ago: Chinese labs trained models almost exclusively on NVIDIA A100/H100
  • 2 years ago: Domestic chips used at small scale for inference
  • 1 year ago: Some training workloads on domestic clusters
  • Now: A 50,000-card domestic cluster pre-training a 1.6T MoE from scratch
  • "Owl Alpha" validated that domestic chips can run inference. LongCat-2.0 validates that domestic chips can train foundation models.

    Anonymous Top-3 on OpenRouter

    "Owl Alpha" ran anonymously on OpenRouter for four weeks and reached top-3 global monthly call volume, meaning international developers—many of them Claude Code / Cursor / Cline users—actively chose the model without knowing its origin. The MOPD Agent expert group is specifically tuned for agentic coding, positioning LongCat-2.0 as a Claude substitute at one-fifth the price with comparable performance.

    Pricing Structure

    The three-tier SiliconFlow pricing—with output ~4× more expensive than input—reflects a typical agent-era pricing model that encourages prompt reuse, lets the model do the thinking, and lets the user do the verification. Versus Claude Sonnet 5's $3 / $15 (input/output), LongCat-2.0 is roughly 1/5 the output price and 1/4 the input price.

    Why It Matters

    1. First Chinese model to approach closed-source frontier in AI coding—SWE-bench Pro 59.5 enters the "industrial usability" range. 2. 50,000-card domestic cluster pre-training a 1.6T MoE—a key infrastructure milestone turning "AI sovereignty" into repeatable engineering. 3. MOPD's purpose-grouped expert routing—a new MoE paradigm likely to be adopted by other Chinese labs. 4. Top-3 global OpenRouter call volume as anonymous preview—real, non-subsidized international adoption. 5. Day-0 launch on SiliconFlow + open source—the first time a Chinese frontier model has shipped simultaneously via open source and an international cloud service.

    Risks and Open Questions

  • Specific hardware configuration of the 50,000-card cluster is undisclosed; whether long-context (1M) training efficiency matches NVIDIA-based clusters needs independent benchmarks.
  • MOPD's "automatic task routing" accuracy in real-world use is untested; ambiguous task boundaries could cause routing errors.
  • SWE-bench Pro 59.5 trails Claude Sonnet 5's 63.2 by ~4 points, which may show up as instability on long-tail agentic coding tasks.
  • OpenRouter "top-3" metric: by call volume or API revenue? The two measures diverge significantly.
  • Real inference cost depends not only on per-token active parameters but also on total parameter loading, expert routing overhead, and cache hit rate—SiliconFlow throughput data will clarify this.
  • References

  • SiliconFlow official post: https://x.com/SiliconFlowAI/status/2071831773076746715
  • China Securities Journal: https://www.cs.com.cn/ssgs/01/2026/06/30/detail_2026063010021451.html
  • Chinaz: https://www.chinaz.com/ainews/29259.shtml
  • IT Home (QQ): https://news.qq.com/rain/a/20260630A0420F00

Tags

#longcat-2.0#meituan#moe#ai-coding#domestic-gpu#swe-bench#openrouter#siliconflow

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208351