How to Self-Host LongCat-2.0 with SGLang

This independent guide summarizes the documented SGLang deployment flow for LongCat-2.0, based on the upstream SGLang pull request that adds LongCat-2.0 inference support and the publisher's GitHub repository. The commands and settings below are taken from the cited sources, and this site does not claim to have independently run this deployment.

Before you begin

The cited SGLang pull request documents a GPU inference path for LongCat-2.0 using the open FP8 weights. The publisher's GitHub README links this path alongside a separate NPU inference path for domestic AI accelerators.

Self-hosting a model of this scale requires substantial infrastructure. The cited PR recommends 16x H20 GPUs using both tensor parallelism and expert parallelism. Confirm the exact hardware requirements and supported versions in the cited PR before provisioning.

• SGLang installed from a version that includes PR #30042 (the cited PR adds LongCat-2.0 support).
• The FP8 model weights, served under the model identifier meituan-longcat/LongCat-2.0-FP8.
• A GPU cluster with 16x H20 GPUs (recommended configuration in the cited PR) for the GPU path, or a domestic NPU environment for the SGLang-FluentLLM NPU path.

This site does not operate a LongCat-2.0 deployment and cannot verify that the cited command succeeds in any particular environment. The cited PR and repository are the source of truth.

Documented GPU launch command

The cited SGLang pull request shows a launch command that starts an SGLang server for the FP8 LongCat-2.0 weights. The command uses --trust-remote-code (required for the custom model code), tensor parallelism of 16 (--tp 16), expert parallelism of 16 (--ep 16), and a concurrency cap of 64 in-flight requests.

• --model points to the FP8 weights under the meituan-longcat HuggingFace organization.
• --tp 16 and --ep 16 combine tensor parallelism and expert parallelism across 16 GPUs.
• --max-running-requests 64 caps concurrent in-flight requests as shown in the cited command.

python -m sglang.launch_server \
  --model meituan-longcat/LongCat-2.0-FP8 \
  --trust-remote-code \
  --tp 16 --ep 16 \
  --max-running-requests 64

Do not remove --trust-remote-code unless the cited documentation states it is safe to do so; custom model code may be required by the LongCat-2.0 architecture.

Hardware recommendation

The cited pull request lists a recommended configuration of 16x H20 GPUs with tensor and expert parallelism. H20 is a high-memory GPU well suited to serving a large MoE model with a 1M-token context window.

This guide does not restate exact VRAM or bandwidth figures, because the cited PR is the authoritative source and hardware guidance may evolve as the upstream integration matures.

NPU inference path

For domestic AI accelerators, the publisher maintains SGLang-FluentLLM with an NPU branch (meituan-longcat/SGLang-FluentLLM, tree npu). The LongCat-2.0 repository links this fork as the NPU deployment option.

Follow the commands and setup steps in the cited NPU branch README. The GPU launch command above does not apply to NPU environments; use the fork-specific instructions instead.

• Clone the publisher's SGLang-FluentLLM repository and check out the npu branch.
• Follow the fork's own setup and launch instructions for the target NPU platform.
• The FP8 GPU command in this guide is not the documented command for NPU deployment.

After the server starts

Once the SGLang server is running, clients can reach it through the OpenAI-compatible interface SGLang exposes on the local host port, the same way any SGLang-served model is called.

Point your client at the SGLang server's base URL instead of the LongCat API platform base URL. The model ID on the SGLang server follows the model identifier used at launch (meituan-longcat/LongCat-2.0-FP8), not the platform identifier.

The exact endpoint paths and request format follow SGLang's standard OpenAI-compatible server behavior. Confirm current behavior in the cited PR and SGLang documentation.

Limitations of this guide

This guide summarizes the deployment path that the cited PR and publisher repository document. It does not add, extrapolate, or recommend settings beyond what the cited sources state.

This site has not benchmarked LongCat-2.0 under SGLang, has not compared it against vLLM or other serving stacks, and has not verified throughput, latency, or memory behavior. The cited PR is the source of truth for supported configurations.

• vLLM and HuggingFace Transformers are documented as alternative serving paths by the publisher; see the LongCat-2.0 model profile for those references.
• Serving performance depends on hardware, batching, and workload; no claims about real-world performance are made here.

Related pages

The pages linked below are independent third-party summaries on this site. They reuse the same source register as this guide and do not introduce new facts.

• LongCat-2.0 model profile

  Verified source links, deployment options, and benchmark table for LongCat-2.0.

• Tools directory

  The SGLang entry with PR #30042 and the SGLang-FluentLLM NPU branch links.

• LongCat API Quick Start

  The hosted API path for developers who prefer not to self-host.

Independent third-party note

This guide is published by an independent third-party site. It is not affiliated with, endorsed by, sponsored by, or operated by Meituan, LongCat, the SGLang project, or any of their affiliates.

No independent deployment or testing has been published by this site. The cited pull request and repository are the source of truth for current commands, model identifiers, and supported configurations.

Self-hosting large open-weight models carries real infrastructure and security responsibilities. Review the cited documentation and your own environment before deployment.

FAQ

The questions below are limited to topics that the cited sources directly support.

Which model identifier does the documented SGLang command use?

The cited SGLang PR launches the FP8 weights under meituan-longcat/LongCat-2.0-FP8 with --trust-remote-code.

What hardware does the cited SGLang configuration require?

The cited PR recommends 16x H20 GPUs using tensor parallelism (--tp 16) and expert parallelism (--ep 16).

Is there a documented NPU deployment path?

Yes. The publisher maintains SGLang-FluentLLM with an NPU branch (meituan-longcat/SGLang-FluentLLM, tree npu) for domestic AI accelerators.

Related model pages

• LongCat-2.0

  A model documented in the meituan-longcat GitHub organization and exposed through the LongCat publisher API platform.

Sources

• SGLang PR #30042 — LongCat-2.0 inference support

  Primary source · Accessed 2026-08-07

  Upstream SGLang pull request adding LongCat-2.0 GPU inference support. Contains the launch command, the FP8 model identifier, and the tensor-parallelism / expert-parallelism settings summarized in this guide.

• LongCat-2.0 GitHub Repository

  Primary source · Accessed 2026-08-07

  Primary repository whose README links the SGLang GPU deployment path and the SGLang-FluentLLM NPU deployment path referenced in this guide.

• meituan-longcat/SGLang-FluentLLM (NPU branch)

  Primary source · Accessed 2026-08-07

  Publisher-maintained SGLang fork with an NPU branch providing optimized inference on domestic AI accelerators.