Model
Model explorer

WizardLM-2 8x22B

OPEN
Microsoft · WizardLM family · released Apr 15, 2024

Evol-Instruct fine-tune of Mixtral 8x22B (141B/~39B active); first open LLM to top 9.0 MT-Bench, pulled by Microsoft ~1 day after release for missing toxicity testing.

ReasoningCodingVisionFunction callingTool useAgentic
251.3
Elo · rank #364
Parameters
141B
Active params
39B (MoE)
Context
64K tokens
Architecture
Sparse Mixture-of-Experts, 8 experts x22B (Mixtral 8x22B backbone)
License
Apache 2.0
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
BIG-Bench HardReasoning48.6%#101
best: ERNIE 4.5 300B-A47B · 94.3%
GPQA DiamondReasoning17.6%#291
best: GPT-6 Astra · 96.0%
IFEvalReasoning52.7%#130
best: Gemma 4 26B A4B · 98.5%
LogicKorHuman preference8.13#15
best: GPT-4o · 9.33
MATH-500Math22.3%#179
best: GPT-5 · 99.4%
MMLU-ProKnowledge40.0%#151
best: Claude Fable 5 · 91.5%
Run it locally
VRAM @ Q4
80 GB
VRAM @ FP16
282 GB
Fits on (Q4)
M3 Max 128GBM3 Ultra 512GBH200 141GBB200 192GB
Throughput data unavailable.
Quantizations
GGUF · AWQ · GPTQ · EXL2
Fine-tune it
Permissive
QLoRA95.9 GB1× H200 141GB
LoRA300.3 GB2× B200 192GB
Full fine-tune2263.1 GBbeyond 8× B200
QLoRA SFT on ~10k samples ≈ $20.17 (1× H200 141GB)
Plan a fine-tune →
WizardLM family
Elo progression across releases
WizardLM-2 8x22BApr 2024251.3
Resources
HFalpindale/WizardLM-2-8x22BGHnlpxucan/WizardLM
API price weights · each benchmark row carries its own source badge (see methodology)