Overview
On June 30, 2026 (05:42 UTC), Meituan's LongCat team officially released and open-sourced its flagship model LongCat-2.0. This is a release backed by detailed engineering specifics:
- Architecture: 1.6T total parameters MoE, ~48B average activation, dynamic activation range 33B-56B / token
- Context: native 1M
- Training hardware: 50,000-card domestic compute cluster (pretrained from scratch)
- Core innovations:
- LSA (sparse attention): efficient scaling to 1M context
- Zero-Compute Experts: dynamic activation with no wasted compute
- MOPD (Mixture of Purpose-Driven experts): experts grouped into three categories (Agent / Reasoning / Interaction), routed by task gating
- SWE-bench Pro: 59.5, described as "comparable to mainstream closed-source models"
- Pricing (SiliconFlow Day 0):
- Input Cache: $0.015 / M tokens
- Input: $0.75 / M tokens
- Output: $2.95 / M tokens
- Preview track record: released anonymously as "Owl Alpha" on OpenRouter, it reached top three globally by monthly call volume; strong performance in Hermes, Claude Code, and similar ecosystems
- Day 0 launch on SiliconFlow for the official release
- 2023: LongCat team began adapting domestic compute, starting from a thousand cards
- Over three years: progressively solved operator adaptation, communication optimization, and training stability
- 2026-06-29: Owl Alpha preview anonymously hit top-three global call volume on OpenRouter
- 2026-06-30: official LongCat-2.0 release + open source
- 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: https://news.qq.com/rain/a/20260630A0420F00
- Dense small models (Qwen 3 series, GLM series)
- Sparse large models (DeepSeek V3, Qwen 3.6 Max)
- Agent group: tool calls, code edits, shell operations
- Reasoning group: math, logic, multi-step reasoning
- Interaction group: conversation, writing, UI interaction
- 3 years ago: Chinese labs relied on NVIDIA A100/H100 for training
- 2 years ago: small-scale domestic chips for inference
- 1 year ago: some training workloads on domestic clusters
- Now: a 50,000-card domestic cluster pretrains a 1.6T MoE from scratch
- Input Cache: $0.015 (very low, incentivizes caching)
- Input: $0.75
- Output: $2.95 (~4x input)
- The exact hardware configuration of the 50,000-card domestic cluster is not fully public; whether it matches NVIDIA-cluster efficiency for 1M-context training needs third-party benchmarks.
- Whether MOPD's task routing accurately auto-classifies in real usage — blurred task boundaries could cause routing errors.
- SWE-bench Pro 59.5 still trails Claude Sonnet 5's 63.2% by about 4 points — potentially showing up as long-tail task stability in real agentic workflows.
- The metric behind the OpenRouter top-three claim (call volume vs API revenue) makes a big difference.
- Real inference cost of 1.6T MoE — total parameter loading, expert routing overhead, and cache hit rates affect deployment economics; SiliconFlow throughput data will be telling.
Timeline:
References:
Analysis
The significance of LongCat-2.0 goes beyond the model itself: it moves Chinese LLMs from "catching up" to "competing on equal footing."
1. 1.6T total / 48B average activation — the MoE route fully mastered by Chinese labs
Chinese LLM development previously split into two camps:
LongCat-2.0 takes the second route and pushes to 1.6T total parameters — first-tier globally. The 33B-56B dynamic activation range is a clever design: simple tasks run at ~33B, complex tasks scale up to 56B, so per-token compute cost tracks task difficulty, avoiding waste on easy inputs.
2. MOPD: turning "experts" from a technical concept into a product feature
MOPD groups experts by purpose:
Combined with Zero-Compute Experts, the engineering benefit is clear: 1.6T total parameters, but per-inference compute cost close to a 48B dense model. This is the economic foundation of the MoE route.
3. 50,000-card domestic cluster — AI sovereignty as engineering reality
Meituan did not disclose the exact chips used, but given the team's domestic-compute adaptation path since 2023, it is likely some combination of Huawei Ascend, Cambricon, and Enflame. The key progression:
Unlike the earlier Owl Alpha stage (fine-tuning on domestic clusters), LongCat-2.0 proves domestic chips can pretrain frontier-scale models, not just run them.
4. Top three on OpenRouter — international users voting with their feet
Owl Alpha ran anonymously on OpenRouter for four weeks and entered the global top three by monthly call volume — meaning international developers (many Claude Code / Cursor / Cline users) actively chose it without knowing who made it. With its Agent-group experts tuned for agentic coding, users can use LongCat-2.0 as a Claude substitute in Claude Code / Cursor at roughly 1/5 of Claude Sonnet 5's standard price with near-comparable performance.
5. Pricing structure logic
This is typical "agent-era pricing": reuse prompts, let the model think and users verify, rather than re-sending long contexts. Against Claude Sonnet 5's standard pricing ($3 / $15), LongCat-2.0's output is 1/5 the price and input 1/4.
Why It Matters
1. A Chinese LLM approaching closed-source frontier performance on AI coding benchmarks — SWE-bench Pro 59.5 is in the "industrially usable" range. 2. 1.6T MoE pretrained on a 50,000-card domestic cluster — a key infrastructure milestone; AI sovereignty becomes repeatable engineering. 3. MOPD's purpose-driven expert grouping — a new MoE engineering paradigm that other Chinese labs may follow. 4. Top-three OpenRouter preview — genuine, non-subsidized international usage. 5. Day 0 open source + international cloud launch — commercialization and ecosystem building in parallel.
Risks and Open Questions
*Note: this is a community forum post; figures and claims are as reported by the author.*