LongCat-Flash-Chat: Meituan's 560B MoE Model Reshaping Enterprise AI

LongCat-Flash-Chat is a huge AI model made by Meituan, with 560 billion parts inside. It's super fast, can handle really long texts, and costs less to use than most rivals. Businesses are already using it for things like better delivery routes and answering customer questions faster. The model is open-source, so anyone can try it, and it could change how companies use AI in the future.

What is LongCat-Flash-Chat and why is it significant for enterprise AI?

LongCat-Flash-Chat is a 560-billion-parameter open-source Mixture-of-Experts (MoE) language model by Meituan, offering faster inference, long context handling, and low-cost API access. It powers real deployments in logistics and enterprise SaaS, reshaping large-scale AI applications and pricing in 2025.

In September 2025, Meituan quietly flipped the switch on LongCat-Flash-Chat , a 560-billion-parameter Mixture-of-Experts (MoE) language model that is already reshaping how large-scale AI is built, priced, and deployed. Below is what practitioners, investors, and researchers are watching - without hype, just the numbers and design choices that matter.

1. A quick anatomy of the model

Attribute	Value / Range	What it means in practice
Total parameters	560 B	Largest Chinese open-source model to date
Active parameters/token	18.6 B - 31.3 B (avg ≈ 27 B)	Cheaper inference than most 100 B+ dense models
Pre-training corpus	~20 T tokens	Comparable to GPT-4-scale data volume
Context length (fine-tune)	32 k → 128 k tokens	Enables long doc QA or multi-turn agent loops
Inference speed	>100 tok/s on H800	Roughly 3× faster than many 70 B dense models
API price floor (public)	$0.70 per 1 M tokens	Undercuts several commercial tier-1 endpoints

Sources:
- arXiv technical report (§3.4 & §5.2)
- OpenSourceForU coverage

2. MoE tricks that actually move the needle

Instead of a generic "throw more GPUs at it" approach, the engineering team baked in four concrete optimizations:

1. Per-layer sub-block
   Two attention blocks + FFN + MoE gate in every layer keeps tensor-parallel communication patterns simple.

2. Zero-Compute "sink" expert
   Tokens scoring below a routing threshold skip heavy computation entirely, shaving ~8 % off average latency.

3. dsv3-like load-bias
   A lightweight bias term prevents the classic "expert 0" hot-spot without extra all-reduce traffic.

4. Inter-layer cross-channel pathways
   Overlaps MoE all-to-all with attention matmuls, cutting bubble time during both training and inference.

3. Benchmark snapshot (late-2025 runs)

Benchmark	Score	Peer comparison note
TerminalBench (agent)	39.5	Ties DeepSeek-Prover-7B-preview on math-heavy turns
τ²-Bench	67.7	+2.4 pts over Qwen3-72B-Instruct
Safety suite avg.	87 %	83 - 94 % across 4 categories

Raw numbers are from an internal eval deck reproduced in the SCMP write-up.

4. Real deployments already live

• Meituan logistics stack - Route-planning agents shaved 11 % off average delivery time in the first 6 weeks of pilot (Futunn news wire, 1 Sep 2025).
• Enterprise SaaS connectors - At least 3 Chinese CRM vendors have rolled LongCat into ticket deflection bots, citing the $0.70/million token rate as the decisive factor.

5. How to access or reproduce

LongCat-Flash-Chat is Apache-2.0 licensed and distributed through:

• GitHub: github.com/meituan-longcat/LongCat-Flash-Chat
• Hugging Face: meituan-longcat/LongCat-Flash-Chat
• Docker images: longcat.ai/inference:latest (includes SGLang backend)

Weights are BF16 shards totaling 1.1 TB; an 8×H800 node loads in ~9 minutes. A quantized INT8 variant drops VRAM usage to 320 GB without measurable accuracy loss on the reported benchmarks.

6. What to watch next

The team's roadmap - outlined in the same arXiv report - lists:

• 256 k context fine-tune (no YaRN)
• Tool-calling grammar compiler (targeting OpenAI-compatible endpoints)
• Community LoRA hub under discussion

If you are benchmarking MoE models for production, LongCat-Flash-Chat now sits at the intersection of lowest publicly documented dollar-per-token cost and top-quartile agent-task performance.

---

What makes LongCat-Flash-Chat special compared with other 560 B models?

Its Mixture-of-Experts (MoE) design keeps only 18.6 B - 31.3 B parameters active per token (average ~27 B) while storing 560 B in total. That is 5 - 15× less active compute than a dense model of the same size yet it still hits competitive scores: 39.5 on TerminalBench and 67.7 on τ²-Bench. In practical terms, inference costs drop to ≈ $0.70 per million tokens - a price point few open-source models at this scale have reached.

Is the model really open-source for commercial use?

Yes. Meituan published the weights, tokenizer, config files, and a detailed 70-page technical report under a permissive license. You can download everything from GitHub, Hugging Face, or the official site longcat.ai without registration fees or usage restrictions.

How fast is inference in production?

Official benchmarks show >100 tokens / s on a single H800 GPU with >90 % speculative acceptance. At that speed a 2 000-token chat turn streams back in under 20 seconds on commodity hardware.

Can it handle long documents?

The context window extends to 32 k tokens out-of-the-box and 128 k tokens after a light continued-pre-training step on ~100 B tokens - no YaRN or other tricks required. Early adopters report accurate summarisation of 80-page PDFs in one pass.

What real-world tasks is it already solving inside Meituan?

Inside Meituan's own stack the model powers:

• Logistics route planning - optimising millions of delivery paths nightly
• Customer-support agents - handling 40 % of chat volume with higher CSAT than the previous pipeline
• Code generation - internal surveys show 28 % faster MR merge times when developers use the built-in coding assistant

These workloads run on the same open weights, proving the efficiency claims are not just lab numbers.