Skip to content

How much VRAM does Mistral Medium 3.5 128B need?

Mistral Medium 3.5 128B needs 72.1 GB for its weights at 4-bit, 75.9 GB with an 8K context and 84.2 GB with 32K.

Parameters
128B
Longest context
256K
Released
Mar 2026
License
Mistral Modified MIT
Strongest at
Chat, Coding, Agents, Docs

Memory at each precision

Memory Mistral Medium 3.5 128B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
As released (FP8)124 GB128 GB137 GB170 GB
6-bit97.5 GB101 GB110 GB143 GB
5-bit84.6 GB88.4 GB96.7 GB130 GB
4-bit72.1 GB75.9 GB84.2 GB117 GB
3-bit58.1 GB62.0 GB70.2 GB103 GB
2-bit44.6 GB48.4 GB56.7 GB89.7 GB
1.6-bit24.5 GB28.4 GB36.6 GB69.6 GB

One conversation, 16-bit context cache, runtime overhead included.

What runs it

The best way Mistral Medium 3.5 128B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst reply
RTX 3060 12 GBToo big
RTX 4060 Ti 16 GBToo big
RTX 3090Too big
RTX 4090Too big
RTX 5090Too big
2× RTX 3090Too big
RTX PRO 6000 Blackwell4-bit52K16 tok/s2.1 s
M4 Max 128 GB5-bit16K3.9 tok/s43 s
M5 Max 128 GB5-bit16K4.3 tok/s13 s
Ryzen AI Max+ 395 128 GB5-bit16K1.7 tok/s25 s
DGX Spark 128 GB6-bit44K1.8 tok/s8.7 s
M3 Ultra 256 GBFP8160K3.9 tok/s24 s
M5 Ultra 512 GBFP8256K5.7 tok/s7.6 s
H100 SXM3-bit48K35 tok/s1.7 s
DGX Station GB300FP8256K34 tok/s0.7 s
8× H200 NVLFP8256K24 tok/s0.3 s

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

Ollama
ollama run mistral-medium-3.5:128b
llama.cpp
llama-server -hf unsloth/Mistral-Medium-3.5-128B-GGUF
vLLM
vllm serve mistralai/Mistral-Medium-3.5-128B
MLX (Mac)
mlx_lm.chat --model mlx-community/Mistral-Medium-3.5-128B-4bit

Questions

How much VRAM does Mistral Medium 3.5 128B need?

Mistral Medium 3.5 128B needs 72.1 GB for its weights at 4-bit, 75.9 GB with an 8K context and 84.2 GB with 32K.

Can Mistral Medium 3.5 128B run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the RTX PRO 6000 Blackwell.

More models