Mixtral 8x22B Instruct V0.1

mistralai · released 2024-04-17 · apache-2.0 license

Mixtral-8x22B-Instruct is Mistral's April 2024 open-weight MoE (141B total / 39B active, Apache-2.0). Mistral's announcement reports strong GSM8K/Math but its standard benchmarks are published only as chart images, so no comparable per-metric numbers could be sourced.

Key specs

TypeLocal open-weight
Parameters140.63B total · MoE, — active
Architecturemixtral
Context window66K tokens
Knowledge cutoff2024-01-31
Modalitiestext
Recommended backends
Minimum viable rig

Benchmark scores

GPQA Diamond
SWE-bench Verified
AIME
MMLU-Pro
BFCL v3 (tool use)
Composite score
Community ratingNo reviews yet

VRAM & disk per quantization

QuantVRAMDiskRAMContext
Q4_K_M83.1 GB81.6 GB66K

Strengths & weaknesses

Strengths: Sparse MoE (141B total / 39B active) with strong math/coding for its 2024 generation; Native function calling, multilingual; Apache-2.0, 64K context

Weaknesses: 2024-era model; not evaluated on most modern benchmarks; Outclassed by current open models