TechCompare LogoTechCompare

How much VRAM does Mistral Medium 3.5 128B need at Q4_K_M? Mistral's open flagship

Mistral Medium 3.5 is a two-card model at Q4: ~150 GB at 256K context, or ~114 GB at 128K on a single 141 GB H200. Dense means no offload cheats, but also no per-token routing penalties.

Mistral Medium 3.5 128B at Q4_K_M needs about 150 GB of VRAM at its native 256K context, so a single 192 GB MI300X or two 80 GB cards run it fully resident. Mistral's July 2026 open-weights release finally gives the Medium line a downloadable model, dense at 128B with no expert routing.

By TechCompare · Updated

Total VRAM required
150 GB
Mistral Medium 3.5 128B at Q4_K_M
Weights
71.7 GB
128B params
KV cache
64.4 GB
256K tokens, FP16 KV

Calculator

Estimated VRAM required

150 GB

128B params at Q4_K_M, 262,144 token context, batch 1, inference.

Weights
71.7 GB
KV cache
64.4 GB
Overhead
13.6 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

MI300X
Datacenter
192 GB
78% used
NVIDIA B300
Datacenter
288 GB
52% used

Just barely too small

Apple M3 Ultra 192GB
Unified
144 GB
short by 5.7 GB
H200 141GB
Datacenter
141 GB
short by 8.7 GB

How this is calculated

Weights at Q4_K_M are 71.7 GB. The KV cache with 8 KV heads at head_dim 128 across 60 layers is 64.4 GB at the full 256K window, plus 13.6 GB of overhead. Being dense, there's no expert-offload trick - every parameter must stay resident, so the floor is set at 256K context by the cache. Capping context to 128K halves the cache to 32 GB and the total to ~114 GB, which fits a single 141 GB H200.

Verdict

The 150 GB total at 256K comes from 71.7 GB of Q4 weights plus 64.4 GB of FP16 KV cache plus 13.6 GB of overhead. Dense architecture means the usual MoE active-only trick doesn't exist here, but inference is also simpler: no cold-expert stalls and predictable throughput. Versus Qwen3.5 122B at the same footprint class, Medium 3.5 trades a slightly larger cache at 256K for Mistral's agentic tuning and European data-residency story. A single MI300X 192GB or H200 pair handles it at full window.

More Mistral 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 Mistral Medium 3.5 open weights?
Yes - the 3.5 release finally made the Medium line downloadable, under Mistral's standard open license. Earlier Medium versions were API-only.
Mistral Medium 3.5 or Qwen3.5 122B?
Both land in the ~100-150 GB class at Q4. Qwen3.5 122B is MoE with a 104 GB footprint and has active-only offload as an escape hatch. Medium 3.5 is dense and simpler to serve with predictable per-token latency, and the Mistral stack integrates more cleanly with European hosting stacks. Benchmark both on your workload - they're close enough that ecosystem decides.