TechCompare LogoTechCompare

How much VRAM does Meituan LongCat-2.0 1.7T need at Q4_K_M? The agentic-coding MoE

LongCat-2.0 at Q4 needs ~1350 GB resident or ~332 GB active-only for 1M context - comparable to DeepSeek V4 Pro's footprint but tuned for agentic coding. Self-host when you need an open agentic loop with no rate limits; otherwise Meituan's API pricing does the job.

Meituan LongCat-2.0 1.7T at Q4_K_M with native 1M context needs about 1350 GB of VRAM with all experts resident, or 332 GB with active-expert offload. Meituan's July 2026 open release activates 48B parameters per token and is tuned specifically for agentic coding loops, making it the most direct open challenger to Kimi K3 in that niche.

By TechCompare · Updated

Total VRAM required
1350 GB
Meituan LongCat-2.0 1.7T (MoE) at Q4_K_M
Weights
952 GB
1700B params
KV cache
275 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

1350 GB

1700B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
952 GB
KV cache
275 GB
Overhead
123 GB

Estimate accuracy: Weights within ~2%. KV cache within ~5% for standard GQA models, ~10% for MLA (DeepSeek). Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.

Hardware that fits

No single GPU in our catalog has enough memory. Multi-GPU or CPU offload required.

How this is calculated

The 1.7T pool at Q4_K_M is 952 GB of weights. The key-value cache with 8 KV heads at head_dim 128 across 64 layers runs 275 GB at the full 1M context, and overhead lands around 123 GB. Active-only loading drops resident weights to 26.9 GB while the cache stays, totaling ~332 GB, which fits four H200 141GB cards. The agentic tuning means the model holds up in multi-hour tool-call sessions where cheaper frontier MoEs drift.

Verdict

The 1350 GB resident budget is 952 GB of Q4 weights, 275 GB of FP16 KV cache at 1M, and 123 GB of overhead - 17 H100 80GB or 10 H200 141GB cards resident. Active-only cuts weights to 26.9 GB and the total to ~332 GB on five H200s. The realistic comparison is Kimi K3 2.8T: K3's hybrid KDA+MLA gives it a much smaller long-context cache, so for long agentic sessions at 1M context K3 is cheaper to hold in memory, while LongCat-2.0 has the edge in agentic-coding evals and is roughly a trillion parameters smaller on the weight side.

More LongCat scenarios

DeepSeek V4 Pro 1.6T (MoE) at Q4_K_M
DeepSeek V4 Pro 1.6T at Q4_K_M with the full 1M-token context needs about 1012 GB of VRAM with every expert resident - this is the real shape of the model and the number to plan a deployment against.
View details ➜
Llama 4 Scout (17B/109B) at Q4_K_M
Llama 4 Scout at Q4_K_M with its native 10M context needs about 2231 GB of VRAM with all 109B params resident - that's the number you size hardware against.
View details ➜
gpt-oss 20B (MoE) at Q4_K_M
gpt-oss 20B at Q4_K_M with native 128K context needs about 19.4 GB of VRAM with all experts resident, dropping to roughly 9.3 GB with active-only weight loading.
View details ➜

Frequently asked questions

What is LongCat-2.0 tuned for?
Agentic coding loops. Meituan trained it to stay coherent across long multi-tool sessions - the pattern where cheaper MoEs start drifting after a few dozen calls. If your workload is chat or batch document work, DeepSeek V4 Pro or GLM-5.3 at similar footprints are cheaper to run.
LongCat-2.0 or Kimi K3 for self-hosted agents?
LongCat-2.0 if the evals favor its agentic tuning on your stack and you can hold ~332 GB active-only. Kimi K3 if long-context memory matters more - its hybrid KDA+MLA attention keeps the 1M-context cache far smaller, though K3 is a bigger pool overall and now carries non-commercial terms.