DEPLOY

Open-source LLM · Mistral AI

Mixtral 8x22B

141 B parameters (39 B active per token) · 64K context · Apache 2.0 · released 2024-04-17 by Mistral AI.

Mixtral 8x22B needs about 282 GB at FP16: no single GPU can hold it. Serving requires an 8-GPU HGX baseboard (INT8/INT4) or a full rack (NVL72, TPU pod) at FP16.

Where this model sits

Mistral's flagship pre-Codestral era. Apache 2.0 with respectable coding + multilingual quality.

What runs it (production serving)

Datacenter GPUs and rack designs sized for Mixtral 8x22B. Consumer cards are not practical at this size.

Memory at different context lengths

Weights are fixed. The KV cache grows linearly with context length and can rival the weights at full window. Most guides only quote weights, then hand-wave a 15% overhead, which is only true at short context.

ContextFP16 totalINT8 totalINT4 total
4K (short chat)282 + 0.9 = 282.9 GB141 + 0.4 = 141.4 GB71 + 0.4 = 70.9 GB
32K (typical RAG)282 + 7.0 = 289.0 GB141 + 3.5 = 144.5 GB71 + 3.5 = 74.0 GB
64K (full window)282 + 14.0 = 296.0 GB141 + 7.0 = 148.0 GB71 + 7.0 = 77.5 GB

KV bytes per token = 2 × 56 layers × 8 KV heads × 128 head-dim × precision bytes. INT4 rows keep the KV cache at INT8 (typical serving default).

Quality benchmarks

Vendor-reported scores from the release blog or tech report. Every row cites its source.

BenchmarkScoreSourceDate
MMLU (5-shot)77.8Mixtral 8x22B blog2024-04-17
HumanEval76.2Mixtral 8x22B blog2024-04-17
MATH (maj@4)41.8Mixtral 8x22B blog2024-04-17

Speed and cost to serve

What cloud providers currently charge to serve Mixtral 8x22B. Tokens per second is what the provider publishes on their own page. Batch size, prompt length, and inference engine can shift these numbers a lot; treat as directional.

ProviderHardwareSpeed$/M in$/M outAs of
Together AIH100 nodes—$1.20$1.202025-08-01
Fireworks AI——$1.20$1.202025-08-01

Cheapest option in this list for output-heavy workloads: Together AI at $1.20/M out. Fastest published: not enough data.

License and variants

License
Apache 2.0
Variants
Base · Instruct
Context window
65,536 tokens (64K)
Architecture
MoE (141B total, 39B active per token)

Compare with

Model card: https://huggingface.co/mistralai/Mixtral-8x22B-v0.1 · Release: Mistral AI announcement

See also: every open-source LLM we track · all AI chips · rack designs.