LongCat-2.0 Requirements: VRAM, GPU & API Pricing (2026)

Running LongCat-2.0 locally needs roughly 3,800GB in BF16 and about 970GB even at INT4 quantization. Neither configuration fits on a single 8-GPU node — you need a multi-node cluster either way. LongCat-2.0 is a Mixture-of-Experts (MoE) model that Meituan announced on June 30, 2026, and released as full weights (safetensors, BF16/F32) on Hugging Face at meituan-longcat/LongCat-2.0 on July 5, 2026. It has 1.6T total parameters, a native 1M-token context window, and supports up to 128K tokens of output. Its standout feature is that both the weights and the inference code are released under the MIT license — a clear example of "open weights" not meaning "runs on a personal PC."

These figures are third-party estimates, not an official hardware requirements spec from Meituan. In real deployments, KV cache memory grows further with batch size and concurrent connections, so treat the table as a floor, not a ceiling.

LongCat-2.0 is an MoE model, so only about 48B parameters activate per token (a dynamic range of 33B–56B); separately, 135B of the 1.6T total is allocated to N-gram embedding. But because which experts fire changes token by token, local inference has to keep all 1.6T total parameters resident in memory. It's easy to assume a ~48B active-parameter model would be lightweight, but memory requirements are driven by total parameters, not active ones — the same structural fact that applies to other large MoE models like DeepSeek V4 and Kimi K3.

- LongCat Sparse Attention (LSA): a sparse attention mechanism built from three components — streaming-aware, cross-layer, and hierarchical indexing — aimed at efficiently handling the native 1M-token context
- Zero-computation experts: simple tokens like punctuation get routed to experts that perform no computation at all, cutting wasted work and keeping the effective active-parameter range down to 33B–56B
- N-gram embedding (135B): 135B of the 1.6T total parameters are allocated to N-gram-based embeddings, which — like ordinary expert weights — still have to sit in memory

Current pricing is roughly 2.5x the launch price for both input and output. Cached context reads remain free, which keeps costs down for workloads that repeatedly send the same long prompt. The API supports both OpenAI-compatible and Anthropic-compatible endpoints, so existing clients can often connect just by switching the base URL.

LongCat-2.0 can be served as an OpenAI-compatible endpoint with either SGLang or vLLM. With SGLang, the base command is python3 -m sglang.launch_server --model-path "meituan-longcat/LongCat-2.0"; with vLLM it's vllm serve "meituan-longcat/LongCat-2.0". Docker images are available for both. That said, given the roughly 3,800GB (BF16) to 970GB (INT4) total memory requirement noted above, actually running it means configuring tensor and pipeline parallelism across a multi-node cluster. Training itself used roughly 50,000 China-made AI ASICs (Huawei Atlas 950-series, over HCCL) on more than 35T tokens — a scale that carries over to inference-side hardware needs as well.

LongCat-2.0 sits at almost the same total-parameter scale (1.6T) as DeepSeek V4-Pro, and both share an MIT license that lets you place the weights on your own infrastructure and modify or commercialize them freely. But since neither runs on a personal PC or a single server, the practical choice comes down to either building a multi-node GPU cluster to self-host, or using the API. For most SMBs evaluating on-premise LLM deployment, self-hosting a frontier-scale MoE model like LongCat-2.0 rarely pencils out — unless there's a specific requirement to keep sensitive data from leaving your own environment, the API is the realistic option.