TechCompare LogoTechCompare

How much VRAM does Qwen 3.8 2.4T Max need at Q4_K_M? Alibaba's biggest 1M MoE

Qwen 3.8 2.4T Max is a ~1839 GB resident deployment at 1M context, or ~419 GB with active-expert offload on four H200 cards. The bigger gate is the license: Qwen 3.8 Max carries non-commercial terms, so check usage rights before sizing hardware.

Qwen 3.8 2.4T Max at Q4_K_M with native 1M context needs about 1839 GB of VRAM with every expert resident. Alibaba's largest open release activates 95B parameters per token from the 2.4T pool. Active-expert offload drops the resident footprint to roughly 419 GB, but the non-commercial license terms mean most production deployments will reach for Alibaba's hosted API instead.

By TechCompare · Updated

Total VRAM required
1856 GB
Qwen 3.8 2.4T Max (MoE) at Q4_K_M
Weights
1344 GB
2400B params
KV cache
344 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

1856 GB

2400B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
1344 GB
KV cache
344 GB
Overhead
169 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 2.4T parameter pool at Q4_K_M is 1344 GB of weights, the largest single number on this list after Kimi K3. The key-value cache with 8 KV heads at head_dim 128 across 80 layers is 328 GB at the full 1M context, and overhead adds about 167 GB. Active-only loading shrinks resident weights to 53.2 GB while the cache stays, totaling ~419 GB. That fits four 141 GB H200 cards, but the license caps commercial use at modest revenue thresholds.

Verdict

The 1839 GB resident budget splits into 1344 GB of weights, 328 GB of FP16 KV cache at 1M, and 167 GB of overhead - 23 H100 80GB cards or 13 H200 141GB cards resident. Active-only cuts weights to 53.2 GB and the total to 419 GB, which fits four H200s, but the per-token PCIe penalty of cold-expert fetch hurts throughput. The licensing reality matters more than the memory math for production: Qwen 3.8 Max ships with non-commercial terms, so for commercial traffic the Alibaba Cloud API or the fully Apache-2.0 Qwen3.5 line is the supported route.

More Qwen 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

Is Qwen 3.8 Max open for commercial use?
Not fully. Qwen 3.8 Max ships with revenue-share style restrictions that cap commercial deployment, unlike the Apache-2.0 Qwen3.5 line. For production commercial workloads, use the hosted Alibaba Cloud API or stick with Qwen3.5 122B, which is fully permissive.
Should I run Qwen 3.8 2.4T Max or Qwen3.5 122B locally?
Qwen3.5 122B for almost everyone. It fits on a single H200 141GB resident and has clean Apache-2.0 licensing. The 2.4T Max needs 13+ H200s resident and carries license limitations that matter the moment you ship to customers.