Beta
Model familiesReading · ~3 min · 73 words deep

Mistral

French AI lab shipping both open-weight (Mistral 7B, Mixtral 8x7B, Mixtral 8x22B) and closed (Mistral Large) models.

Mistral family on /family
TL;DR

French AI lab shipping both open-weight (Mistral 7B, Mixtral 8x7B, Mixtral 8x22B) and closed (Mistral Large) models.

Level 1

Mistral AI was founded in 2023 in Paris by former Meta and DeepMind researchers. Their first release, Mistral 7B, set a quality-per-size benchmark for open-weight models. Mixtral 8x7B (Dec 2023) introduced MoE to the open ecosystem. Mistral Large is their closed-API flagship. Mistral positions on EU data sovereignty and GDPR-native compliance · a competitive differentiator in European enterprise sales.

Level 2

Mistral architecture: sliding-window attention, RoPE, Grouped-Query Attention. Mixtral uses MoE with 8 experts total, 2 active per token. Mistral Large and subsequent closed models scale further with undisclosed size. Mistral's training infra is primarily EU-based · with European data residency as a key feature. Mistral Small / Medium / Large tiers differentiate on size and price. The lab has raised $500M+ at a $14B+ valuation as of 2026, becoming Europe's AI national champion.

Level 3

Mistral 7B architecture: 32 layers, 4096 hidden, sliding-window attention with 4K window + 8K total context. Mixtral 8x7B: 8 experts of 7B each, 2 active per token, 47B total / 13B active parameters. Mixtral 8x22B: same pattern scaled up. Mistral Large (closed) is believed to be larger MoE, undisclosed specifics. Training data: European-language emphasis, strong in French, German, Spanish, Italian. Distribution: direct API, Azure AI Studio, AWS Bedrock, OCI. EU data residency is default, not optional · legal positioning unique vs Western frontier labs.

Why this matters now

Mistral's EU-native positioning is the European enterprise wedge. EU AI Act and GDPR favor Mistral over US labs for sovereignty-sensitive deployments.

The takeaway for you
If you are a
Researcher
  • ·Sliding-window attention, MoE (Mixtral), GQA
  • ·Mixtral 8x7B was the open-weight MoE breakthrough
  • ·Training on EU infra for data residency
If you are a
Builder
  • ·Mixtral 8x22B for self-hosted high-capability workloads
  • ·Mistral Large via API for closed-quality
  • ·Use Mistral if EU data residency is a hard requirement
If you are a
Investor
  • ·Europe's AI national champion · political/regulatory tailwinds
  • ·$14B valuation lags Anthropic/OpenAI · compelling multiple
  • ·GDPR + EU AI Act compliance is their moat
If you are a
Curious · Normie
  • ·French AI company
  • ·Ships both free-download and paid-API models
  • ·Default choice for European businesses that care about data privacy
Gecko's take

Mistral's bet is on regulatory moat, not raw capability. In a world where EU AI Act compliance matters, that's a valid bet.

Mistral, Mixtral (open-weight variants) yes · Apache 2.0. Mistral Large and other closed models via API with per-token pricing.