Meituan LongCat
Listed at https://github.com/meituan-longcat
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | OpenRouter: $0.60/M token (48B active MoE) | per million tokens |
| Output | OpenRouter: $2.40/M token (48B active MoE) | per million tokens |
🤖 Supported Models (4)
✨ Pros
- ✓1.6T total / 48B active MoE, operational cost close to DeepSeek V3
- ✓1M token context window, +13 point LongBench v2 lead
- ✓MIT license, no commercial restriction
- ✓OpenRouter model meituan/longcat-2-0-1.6t-moe ready to use
- ✓vLLM-compatible inference, single-node 8x H100 INT4 deploy
- ✓Distilled 70B / 140B variants for budget-constrained single/dual-node
⚠️ Cons
- ×No multimodal support (text-only)
- ×No published function-calling benchmarks
- ×INT4 quantization loses ~2-3 MMLU points
- ×Full BF16 model needs multi-node deployment (~3.2 TB VRAM)
- ×Limited Chinese documentation and community examples
- ×Routing requires OpenRouter or self-hosted vLLM cluster
🎯 Best For
Long-context RAG, full-codebase analysis, multi-hour document processing, self-hosted open-weight deployment
💰 Pricing & Plans
| Model | Input ($/M tokens) | Output ($/M tokens) | Best For |
|---|---|---|---|
| LongCat-2.0 (1.6T MoE) | $0.60 | $2.40 | Frontier open-weight MoE via OpenRouter |
| LongCat-2.0 INT4 quantized | $0.30 | $1.20 | Self-host, INT4 cost reduction (~50%) |
| LongCat-2-140B | Self-host only | Self-host only | Mid-size distillation, single-node budget |
| LongCat-2-70B | Self-host only | Self-host only | Smallest open variant, dual-node or single H100 |
| Self-host on RunPod/Lambda/CoreWeave | GPU-hour pricing | GPU-hour pricing | Full control, pay for compute not tokens |
🔧 API & Developer Experience
- •Distribution Channels: OpenRouter pay-as-you-go (meituan/longcat-2-0-1.6t-moe) for hosted inference; Hugging Face weights for self-host; no first-party commercial API.
- •API Compatibility: On OpenRouter the model is OpenAI-compatible — same /chat/completions shape, function-calling via tools parameter, SSE streaming. Self-hosted vLLM or SGLang exposes the same surface.
- •License: MIT-licensed weights on Hugging Face. No commercial restriction clauses, no call-home telemetry, no per-user royalty — drop-in for proprietary products.
- •Context Window: 1M tokens on the flagship; ~60 min video subtitles + 1.5× novel-length text in a single request, comparable to Gemini 1M context tier.
- •Inference Stack: vLLM and SGLang compatible; single-node 8×H100 INT4 deployment for the 1.6T MoE; BF16 needs multi-node (~3.2 TB VRAM).
- •Quantization Options: INT4 ships ready-to-run with ~2-3 MMLU point loss vs BF16; FP8 is the recommended sweet spot for production (better quality / cost ratio).
- •Ecosystem Maturity: Chinese-language docs and community examples lag DeepSeek/Qwen; English tutorials sparse. Most deployments rely on copies of vLLM configs from the official repo.
🐾 Sparse MoE at DeepSeek-V3 Op-Cost
LongCat-2.0 is a 1.6T-total / 48B-active sparse-activated MoE that Meituan explicitly prices to operate at roughly the cost of DeepSeek V3 — about ¥0.30 / ¥1.20 per million tokens in INT4, or $0.60 / $2.40 on OpenRouter pay-as-you-go (BF16-equivalent quality). The headline numbers translate into practical deployment economics: a single 8×H100 INT4 node can serve roughly the same queries per hour as a DeepSeek V3 cluster at one-third the GPU-hour cost, while delivering a +13-point lead on LongBench v2 (the long-context Chinese benchmark that Meituan helped maintain). The trade-off versus a dense model at the same effective size (Qwen3-72B, Llama-3.1-70B) is operational rather than capability: developers need MoE-aware serving (vLLM's `enable_expert_parallelism`, SGLang's `--enable-ep`) and to size KV-cache for active experts rather than total parameters. Meituan publishes the routing config alongside the weights, so the integration work is configuration, not re-engineering — but it is not zero work for teams that have never operated a sparse MoE in production.
🌐 Regional Availability & Latency
LongCat has no first-party commercial API endpoint — access is split between two channels. Inside China, developers typically download the MIT weights from Hugging Face (or mirror sites) and self-host on RunPod-equivalent H100 clusters inside HK / SG / JP regions; domestic GPU-cloud providers (Alibaba Bailian, Tencent Cloud, Volcengine) intermittently list the 1.6T MoE on their model marketplaces. Outside China, OpenRouter serves the model under `meituan/longcat-2-0-1.6t-moe` at $0.60 / $2.40 per million tokens with no mainland routing. Latency depends entirely on the deployment region. Self-hosted clusters in HK / SG deliver 100-300ms first-token to mainland China users (after the Great Firewall penalty, which adds 50-100ms); OpenRouter routed via OpenAI-compatible aggregators (FreeModel, Portkey) typically adds 200-400ms for users in China. For production APIs serving Chinese traffic, the standard recommendation is either (a) self-host on an HK-region H100 cluster, or (b) use a domestic model marketplace — both avoid the cross-border routing penalty entirely.