Skip to main content
MicrosoftMoE

WizardLM-2 8x22B RAM Calculator

For WizardLM-2 8x22B, plan about 128GB system RAM at Q4_K_M / 8K context — MoE still loads ~176B total weights even though only 44B active/token run per token. WizardLM-2 8x22B weights are available for local runtimes (llama.cpp / Ollama / vLLM class stacks) — buy kits you can fill with dual-channel DDR5 (or ECC RDIMM on true workstations).

WizardLM-2 8x22B is Microsoft AI's most advanced Wizard model. It demonstrates highly competitive performance compared to leading proprietary models, and it consistently outperforms all existing state-of-the-art opensource models. It is...

Standard Recommendation

128GB RAM

Calculated for 4-bit (Q4_K_M) @ 8K Context

1. Workload

Inference sizes run-time memory. Training adds optimizer/activation headroom and steers toward ECC.

2. Hardware path

CPU + RAM offload path: full model weights reside in system RAM (llama.cpp / similar). Dual-channel DDR5 bandwidth is the speed bottleneck.

3. Quantization

GGUF-style bit widths for planning. Native FP4/FP8 trainer footprints can differ.

4. Context length

Grows KV cache (inference) or activation scratch (training ballpark).

8,192 tokens

Inference bandwidth snapshot

DDR4 ~45 GB/s

0.5 t/s

DDR5 ~96 GB/s

1.0 t/s

Unified ~300 GB/s

3.0 t/s

VRAM ~1008 GB/s

10.2 t/s

Host RAM target

128GB

Inference · CPU offload · Q4 K_M

64GB96GB128GB192GB256GB
Model weights:99 GB
KV cache:0.11 GB
OS / runtime:8 GB
Host total:107.1 GB

Kit picks (128GB)

Disclosure: As an Amazon Associate I earn from qualifying purchases. Rankings use price and spec data only — not paid placement. How we rank products

A-Tech 128GB Kit (4x32GB) RAM for Apple iMac 2019 & 2020 27 inch Retina 5K | DDR4 2666 MHz SODIMM PC4-21300 / PC4-21333 260-Pin SO-DIMM Max Memory Upgrade

SO-DIMM4-stick kit
$825.02$6.45/GBIn stock

Four sticks can stress the memory controller and lower stable XMP speeds on many consumer boards.

Laptop / mini-PC form factor — will not fit desktop DIMM slots.

DetailsBuy on Amazon →

A-Tech 128GB Kit (4x32GB) DDR5 4800MHz PC5-38400 CL40 SODIMM 2Rx8 Dual Rank 1.1V Non-ECC Unbuffered SO-DIMM 262-Pin Laptop Computer RAM Memory Upgrade Modules

SO-DIMMECC4-stick kit
$1402.16$10.95/GBIn stock

Four sticks can stress the memory controller and lower stable XMP speeds on many consumer boards.

Laptop / mini-PC form factor — will not fit desktop DIMM slots.

DetailsBuy on Amazon →

Corsair Vengeance LPX 128GB (4x32GB) DDR4 3600 (PC4-28800) C18 1.35V Desktop Memory - Black

UDIMM4-stick kit
$194.99$1.52/GBIn stock

Four sticks can stress the memory controller and lower stable XMP speeds on many consumer boards.

Confirm motherboard QVL / max capacity per slot before buying.

DetailsBuy on Amazon →

NVTEK 128GB (4X32GB) DDR4 2666MHz PC4-21300 UDIMM 2Rx8 1.2V CL19 288-PIN Non-ECC Unbuffered Desktop PC Computer Memory KIT

UDIMMECC4-stick kit
$794.49$6.21/GBIn stock

Four sticks can stress the memory controller and lower stable XMP speeds on many consumer boards.

Confirm motherboard QVL / max capacity per slot before buying.

DetailsBuy on Amazon →

A-Tech 128GB Kit (4x32GB) DDR5 4800MHz PC5-38400 CL40 UDIMM 2Rx8 Dual Rank 1.1V Non-ECC Unbuffered DIMM 288-Pin Desktop PC/Computer RAM Memory Upgrade Modules

UDIMMECC4-stick kit
$1728.68$13.51/GBIn stock

Four sticks can stress the memory controller and lower stable XMP speeds on many consumer boards.

Confirm motherboard QVL / max capacity per slot before buying.

DetailsBuy on Amazon →

NEMIX RAM 128GB (4X32GB) DDR4 3200MHz PC4-25600 2Rx8 1.2V CL22 288-PIN ECC Unbuffered UDIMM KIT

UDIMMECC4-stick kit
$1193.49$9.32/GBIn stock

Four sticks can stress the memory controller and lower stable XMP speeds on many consumer boards.

Confirm motherboard QVL / max capacity per slot before buying.

DetailsBuy on Amazon →
All 128GB prices →Check board fit in RAM Finder →

Why WizardLM-2 8x22B pressures system RAM

WizardLM-2 8x22B is Mixture-of-Experts: inference activates 44B active/token, but VRAM/RAM must usually hold the full ~176B expert set for fast routing. At Q4 the weight slab is ~99GB before KV (~0.11GB at 8K) and ~8GB OS/runtime overhead — totaling ~107.1GB raw, rounded to a 128GB kit. Stretching toward the full 66K-token window multiplies KV far faster than weights; that is the usual “I bought enough RAM for the model but still OOM” failure on Microsoft MoE pages.

What RAM kit to buy

Buy a matched dual-channel DDR5 kit at 128GB for WizardLM-2 8x22B (EXPO/XMP only if stable). Avoid single-stick installs — local inference is bandwidth-sensitive when layers spill to host memory. Pair with Apple Mac Studio (192GB Unified Memory) or Institutional Node (8x H100 / A100) when staying in the Enterprise GPU Node / Mac Studio 192GB tier, and keep 20–30% RAM free for the OS + browser.

Workload notes

Microsoft/Phi-class models like WizardLM-2 8x22B target efficient local assistants — do not overspend on 256GB+ kits unless you stack multiple sessions. At 176B, WizardLM-2 8x22B sits in the large local-LLM band: Q4 on a strong GPU is realistic, FP16 usually is not on consumer cards. Release window noted as 2025/2026; always re-check the model card before buying hardware for a specific checkpoint.

Next steps:128GB RAM pricesDDR5 RAM pricesCapacity comparisonRAM Finder

Technical Specifications

Total Parameter Count176 Billion
Active Parameters Per Token44 Billion
Maximum Context Window66K tokens
Primary Framework SupportOllama, llama.cpp, ExLlamaV2, vLLM

GPU & VRAM Sizing Profile

Enterprise GPU Node / Mac Studio 192GB
Est. VRAM Required111 GB VRAM
Target GPU HardwareApple Mac Studio (192GB Unified Memory) or Institutional Node (8x H100 / A100)

Hardware Profile: Server-scale deployment. Running this model locally requires extreme unified memory Apple systems or professional multi-GPU servers.

WizardLM-2 8x22B Memory FAQs

How much RAM for WizardLM-2 8x22B at Q4 vs FP16?

At Q4_K_M with an 8K context we estimate ~128GB system kits for WizardLM-2 8x22B (weights ~99GB). FP16 jumps to roughly a 384GB kit class and often wants 111GB-class VRAM instead of host RAM alone — use the on-page calculator to retarget context and quant.

Does MoE mean I only need RAM for 44B active params on WizardLM-2 8x22B?

No. WizardLM-2 8x22B still stages ~176B total expert weights for fast routing even though only 44B active/token compute each token. Size RAM/VRAM from total parameters (and KV), not active-only marketing figures.

What GPU tier fits WizardLM-2 8x22B?

Enterprise GPU Node / Mac Studio 192GB: target about 111GB VRAM (Apple Mac Studio (192GB Unified Memory) or Institutional Node (8x H100 / A100)). Server-scale deployment. Running this model locally requires extreme unified memory Apple systems or professional multi-GPU servers.

Can I run WizardLM-2 8x22B with less than 128GB if I lower context?

Yes — shorter context shrinks KV (~0.11GB at 8K). Dropping to 2K–4K context can fit smaller kits, but keep OS headroom; paging kills tokens/s more than a slightly larger kit costs.

Same VRAM tier

Models that land in the same hardware profile (Enterprise GPU Node / Mac Studio 192GB) at Q4 / 8K context.

Mixtral 8x22B InstructQwen3 VL 235B A22B ThinkingQwen3 VL 235B A22B InstructQwen3 235B A22B Thinking 2507Qwen3 235B A22B Instruct 2507Qwen3 235B A22B

Related Models

Phi 4Phi-4-mini (3.8B)Mixtral 8x22B InstructQwen3 VL 235B A22B ThinkingQwen3 VL 235B A22B InstructQwen3 235B A22B Thinking 2507Qwen3 235B A22B Instruct 2507Qwen3 235B A22B