LongCat-Flash-Lite

Overview

LongCat-Flash-Lite is a 68.5-billion-parameter efficient Mixture-of-Experts (MoE) model released by Meituan's LongCat team in February 2026. Despite its size, it activates only approximately 3 billion parameters per inference — achieving a remarkable 22:1 ratio of total-to-active parameters. This makes it the most computationally efficient model in the LongCat family, designed for scenarios where inference speed and cost are paramount: API serving, edge deployment, and high-throughput agentic workflows. The key enabling technology is an N-gram embedding table that improves parameter utilization by pre-computing token combinations, mitigating the I/O bottleneck that typically limits MoE inference speed. The model supports a 320K output token limit and is available with 50 million free tokens per day through the LongCat API platform. It also serves as the backbone for LongCat-Next, the discrete native multimodal model.

Architecture

Flash-Lite uses the same ScMoE (Shortcut-connected Mixture of Experts) architecture as Flash-Chat, but with fewer total experts and a smaller parameter budget. The N-gram embedding module pre-computes 5-gram token combinations and stores them in a compact lookup table, reducing the number of routing decisions the MoE must make per token. This addresses a fundamental bottleneck in MoE models: expert routing requires reading model parameters from GPU memory, and with many small experts, the I/O cost can dominate the computation cost. By reducing the routing granularity through N-gram pre-computation, Flash-Lite achieves inference throughput closer to a dense model of comparable size while maintaining the parameter capacity of a 68.5B model. The publisher also reports specialized cache and kernel optimizations for the LongCat API platform that further reduce end-to-end latency.

Deployment

Flash-Lite is available through the LongCat API platform with 50 million free tokens per day and a 320K maximum output token limit. Model weights are on HuggingFace (meituan-longcat/LongCat-Flash-Lite) under the MIT License. Due to the compact ~3B activated parameter count, Flash-Lite can be served on substantially fewer GPUs than the 560B models — the publisher reports that it can run on standard instances through Hugging Face Transformers with trust_remote_code=True, making it the most accessible self-hosted LongCat model.

Capabilities

Described in the cited primary documentation. Not an independent performance evaluation.

• 68.5B total parameters (MoE), ~3B activated per inference.
• N-gram embedding for superior parameter utilization.
• Optimized cache and kernel for low-latency inference.
• 320K maximum output tokens via API.
• 50M free tokens/day on LongCat API platform.

Access & License

MIT License.

Documented context window: 320,000 tokens (per the cited publisher API documentation).

Sources

• LongCat-Flash-Lite on HuggingFace

  Primary source · Accessed 2026-07-27

  68.5B total, ~3B activated. N-gram embedding for parameter efficiency.

• LongCat API Changelog — Flash-Lite Release

  Publisher documentation · Published 2026-02-05 · Accessed 2026-07-27

  Confirmed release date, parameter specs, N-gram embedding, and optimization for inference efficiency.

FAQ

Where can the primary source for LongCat-Flash-Lite be found?

The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Flash-Lite.

Is LongCat-Flash-Lite available through the LongCat API?

LongCat-Flash-Lite is documented in the meituan-longcat GitHub organization. Refer to the publisher API documentation at longcat.chat for current API availability.

Is this page an official LongCat or Meituan page?

No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.

How often is this page updated?

This page was last verified on 2026-07-27. Content is reviewed when new publisher documentation or model releases become available.