back to local model checker

Mixtral 8x22B

MoETools

Mistral AI · 141B (39B active) · Mixture of Experts

Large MoE with 39B active params

Apache 2.0
Downloads

16.5K

Likes

746

Released

2024-04

Context

64K tokens

Use cases

chatcodereasoning
Total experts8
Active experts2
Active params39.1B

Your hardware

16 GB Reference Desktop

Quant
Q4_K_M
Memory
72.7 GB
Speed
--
Fit
Too heavy

Quantization options

What is quantization?
QuantBitsVRAMDiskQualityStatus
Q2_K245.6 GB43.1 GBlowToo heavy · ~1 t/s
Q3_K_M363.7 GB60.3 GBmoderateToo heavy · ~1 t/s
Q4_K_M472.7 GB68.9 GBgoodToo heavy · ~1 t/s
Q5_K_M590.8 GB86.2 GBgoodToo heavy · ~1 t/s
Q6_K6108.8 GB103.4 GBexcellentToo heavy · ~1 t/s
Q8_08144.9 GB137.9 GBexcellentToo heavy · ~1 t/s
F1616289.4 GB275.8 GBlosslessToo heavy · ~1 t/s

About this model

Mixtral 8x22B is a 141B sparse mixture-of-experts (MoE) model activating 39B active per token from Mistral AI in the Mistral family, designed primarily for chat, code, reasoning workflows. Supports structured tool calling; use Q4_K_M or higher when strict JSON schema adherence matters. Supports up to 64K context tokens; start at 4K–8K with Q8_0 KV cache on tight GPUs.

Hardware planning baselines start at 78.8 GB minimum memory and 131.3 GB recommended memory. At Q4_K_M, weights plus baseline runtime overhead use roughly 72.7 GB of RAM/VRAM and 68.9 GB of disk space. At 141B, Mixtral 8x22B belongs firmly in the workstation and multi-GPU tier. Running it comfortably requires 131.3 GB of total memory across dual RTX GPUs, high-memory Apple Silicon (Mac Studio Ultra/Max), or MoE hybrid CPU-RAM offloading—mobile devices and standard 16 GB laptops are too constrained for practical speeds.

Deploy Mixtral 8x22B locally via Ollama (`mixtral:8x22b`), Ollama with Flash Attention enabled. Under Apache License 2.0, review official weight cards before commercial redistribution.

Highlights

  • Local Fit Baseline78.8 GB minimum and 131.3 GB recommended RAM/VRAM.
  • Primary QuantQ4_K_M targets 72.7 GB memory and 68.9 GB disk.
  • Context & Capabilities64K max context · Sparse MoE · Tool calling.

System requirements by device

Memory planning targets 78.8 GB minimum RAM/VRAM and 131.3 GB recommended memory. The baseline quant is Q4_K_M, estimated at 72.7 GB memory and 68.9 GB disk. Actual runtime usage shifts with context length, KV cache precision, and OS background overhead.

DeviceNotes
Mac (Apple silicon)Requires a 48 GB–192 GB Mac Studio or MacBook Pro Max/Ultra with enough unified wired budget for 72.7 GB.
Windows / Linux GPUUse dual/multi-GPU CUDA splitting (--tensor-split) or MoE CPU expert offloading (-ot) paired with 32–128 GB DDR5 system RAM.
Phones & Compact Edge BoardsNot practical on phones (78.8 GB min RAM exceeds Android LMK and iOS Jetsam limits). Pick a 2B–4B model on mobile.

Start with a short session and leave memory for the operating system. A file fitting on disk does not guarantee that it will load in memory.

Estimated memory by quantization

These are rough estimates, not tested minimum requirements. Check the actual file and runtime documentation; context, offloading and operating-system overhead can change the memory needed.

Quant TierEstimated memoryDisk SizeQuality
Q2_K45.6 GB43.1 GBlow
Q3_K_M63.7 GB60.3 GBmoderate
Q4_K_M72.7 GB68.9 GBgood
Q5_K_M90.8 GB86.2 GBgood
Q6_K108.8 GB103.4 GBexcellent
Q8_0144.9 GB137.9 GBexcellent
F16289.4 GB275.8 GBlossless

How to run Mixtral 8x22B on your device

Start with a short prompt and the smallest quant that still gives the quality you need. A clean first run is more useful than loading a larger file with no memory headroom.

Windows

Run Mixtral 8x22B on Multi-GPU or High-RAM Desktop

  1. Plan for 78.8 GB minimum combined memory and 131.3 GB recommended memory for Q4_K_M.
  2. Use llama.cpp `-ot ".ffn_.*_exps.=CPU"` or `--tensor-split` across multiple GPUs so attention and KV cache remain in fast VRAM.
  3. Enable Flash Attention (`-fa on`) and Q8_0 KV cache (`-ctk q8_0 -ctv q8_0`) to prevent context buffers from exceeding GPU headroom.
  4. Avoid attempting to run workstation-scale weights on smartphones or 8 GB laptops; use a 2B–8B checkpoint on portable devices.
Device how-to guide
Mac

Run Mixtral 8x22B on High-Memory Apple Silicon Mac

  1. Verify you are on a 48 GB–192 GB Mac Studio or MacBook Pro Max/Ultra capable of holding 72.7 GB in wired unified memory.
  2. Launch via `llama-server` or LM Studio with MLX/Metal enabled and set initial context conservatively to 4K–8K.
  3. Monitor Activity Monitor Memory Pressure to confirm zero SSD swap thrashing during token generation.
  4. Step down to a lighter quantization (e.g., Q3_K_M or IQ4_XS) if total memory utilization exceeds 75%.
Device how-to guide

Mixtral 8x22B vs alternatives

ModelParamsContextMin RAMLicense
Mixtral 8x22B141B (39B active)64K78.8 GBApache 2.0
Devstral 2 123B123B256K68.7 GBMRL
Mistral Small 4 119B119B (6.5B active)256K66.5 GBApache 2.0
Mixtral 8x7B47B (12.9B active)32K26.3 GBApache 2.0
Devstral Small 2 24B24B256K13.4 GBApache 2.0
Mistral Small 3.1 24B24B128K13.4 GBApache 2.0

Use this table to choose the smallest nearby model that still matches the job. A related model with lower RAM can be the better local choice when the current page fits only at the edge of available memory.

Frequently asked questions

What is Mixtral 8x22B?
Mixtral 8x22B is a 141B sparse mixture-of-experts model (39B active active per token) from Mistral AI, supporting up to 64K tokens with 78.8 GB minimum RAM and 131.3 GB recommended RAM.
How much VRAM or RAM does Mixtral 8x22B require?
At Q4_K_M, Mixtral 8x22B occupies roughly 72.7 GB of active memory and 68.9 GB on disk. Add 2–6 GB of spare headroom for KV cache buffers, runtime scratch space, and OS apps (131.3 GB recommended).
What is the best way to run Mixtral 8x22B locally?
Use `ollama run mixtral:8x22b` for a quick CLI/API setup, or load Q4_K_M in LM Studio or llama.cpp with Flash Attention enabled when you want explicit KV cache and layer control.
Can Mixtral 8x22B run on a consumer laptop or smartphone?
No smartphone or 8–16 GB laptop can hold Mixtral 8x22B cleanly in local memory. It targets multi-GPU NVIDIA workstations, high-RAM Mac Studios, or MoE CPU-RAM hybrid rigs with at least 78.8 GB total memory.
How does Mixtral 8x22B compare to Devstral 2 123B?
Devstral 2 123B (123B, 68.7 GB min RAM) sits closest in size. Compare active parameters, context ceiling (64K vs 256K), and license terms (Apache 2.0) before switching.