Meituan LongCat-2.5 (1.78T MoE) Architecture Parameters, Benchmarks & VRAM
Meituan MIT-licensed 1.78T-parameter multimodal agentic MoE with 768 experts.
🧬 Meituan LongCat-2.5 (1.78T MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (meituan-longcat/LongCat-2.5) config.json & Model Card
1. Backbone & Dimensions
pro_v- Total / Active Params
- 1.78T / 48B
- Hidden Size (d_model)
- 8,192
- MoE Experts
- 768 Routed
- Activation / Norm
- SiLU (SwiGLU) · ε=0.00001
- Native Tensor Dtype
- BF16 / FP8
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 38 Layers
- Hidden Layer Layout
- 38-Layer Sparse MoE (768 Experts)
- Attention Heads (Q / KV)
- 64 Q / 8 KV (8:1 GQA)
- Head Dimension
- 128
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 524,288 tokens
- RoPE Configuration
- RoPE (原生 524,288 Context)
- Vocab Size
- 163,840 (Untied)
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 Meituan LongCat-2.5 (1.78T MoE) Full Benchmark Scorecard
19 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 11 items
Terminal-Bench 2.1 is an updated release of the Terminal-Bench benchmark that tests AI agents' ability to operate a computer via the terminal.
MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.
PaperBench is a benchmark for evaluating AI agents on their ability to replicate research papers. It tests models on complex, multi-step workflows involving code implementation, experimentation, and reproducing scientific results from academic publications.
Vision2Web evaluates multimodal models on converting visual designs and screenshots into functional web pages, measuring end-to-end design-to-code capability.
Claw-Eval tests real-world agentic task completion across complex multi-step scenarios, evaluating a model's ability to use tools, navigate environments, and complete end-to-end tasks autonomously.
DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.
SkillsBench evaluates coding agents on self-contained programming tasks, measuring practical engineering skills across diverse software development scenarios.
NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.
SWE Atlas - Codebase QnA evaluates a model's ability to answer questions about real codebases, measuring repository-level comprehension and the ability to reason about code structure, behavior, and intent across an entire project.
Terminal-Bench 4.0 evaluates AI agents on real-world work in terminal and command-line containerized environments across 66 tasks, with emphasis on science-adjacent and frontier engineering problems.
Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation.
🤖 Agents & Tools 4 items
BrowseComp is a benchmark comprising 1,266 questions that challenge AI agents to persistently navigate the internet in search of hard-to-find, entangled information.
JobBench evaluates agents on work-related tasks, as reported in the MiMo-V2.6 release.
AutomationBench v1.0.6 evaluates agent performance on multi-step workflow automation tasks.
SpreadsheetBench 2 evaluates office automation agents on spreadsheet analysis, reasoning, and manipulation tasks.
📐 Math 2 items
👁️ Multimodal 2 items
CharXiv-R is the reasoning component of the CharXiv benchmark, focusing on complex reasoning questions that require synthesizing information across visual chart elements.
Video-MME is the first-ever comprehensive evaluation benchmark of Multi-modal Large Language Models (MLLMs) in video analysis.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ Meituan LongCat-2.5 (1.78T MoE) Interactive VRAM & Token Speed Calculator
Physical 96 GB · Load 1171.2%
Exceeds usable VRAM by 1007.79 GB
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run meituan-longcat-2-5Insufficient Local VRAM? Ways to Run Meituan LongCat-2.5 (1.78T MoE)
Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.
📊 Meituan LongCat-2.5 (1.78T MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 3560 GB | 0.72 GB | 3632.37 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 1891.25 GB | 0.72 GB | 1930.24 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 1459.6 GB | 0.72 GB | 1489.96 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 1266.03 GB | 0.72 GB | 1292.51 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 1079.12 GB | 0.72 GB | 1101.87 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 945.63 GB | 0.72 GB | 965.7 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 869.98 GB | 0.72 GB | 888.54 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 700.88 GB | 0.72 GB | 716.06 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for Meituan LongCat-2.5 (1.78T MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of Meituan LongCat-2.5 (1.78T MoE)?
Meituan LongCat-2.5 (1.78T MoE) uses LongcatCausalLM with 38 layers (38-Layer Sparse MoE (768 Experts)), hidden_size=8192, 64 Q heads / 8 KV heads (head_dim=128), and vocab_size=163,840.
How much VRAM does Meituan LongCat-2.5 (1.78T MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), Meituan LongCat-2.5 (1.78T MoE) requires 1101.87 GB VRAM. Full FP16 requires 3632.37 GB.