AI Models. Mistral AI
Self-host / Europe / multimodal / code

Mistral AI

Medium 3.5, Small 4, Apache 2.0, GPAI Code

Mistral updates its commercial line: Medium 3.5 and Small 4 (Mar 2026) alongside open-weight Mistral 3 (Large 3: MoE 41B active/675B total; smaller dense 3B/8B/14B), Pixtral 12B (VLM, text+image). Medium 3.5 is the 'SOTA' cloud variant; Small 4 is an efficient dense model. Mistral supports the EU GPAI Code.

Verified: 2026-05-22

Purchase decision (when to choose / when to avoid)

Choose if...

  • You want an 'EU-friendly' alternative and self-host/open-weight options (depending on model).
  • You're building solutions on your own infrastructure and want to control data.
  • You need good document/code models at a reasonable cost.

Avoid if...

  • You need a single, always 'best' frontier model without variant selection.
  • You lack MLOps competence for self-host — SaaS/API from top providers will be simpler.

Cost in practice (scenarios)

SMB (copy + documents)

SaaS/Pro/Team or API; usually cheaper than premium top-3.

  • moderate volume
Self-host

Cost is GPU infrastructure + maintenance; pays off at steady volume.

  • MLOps
  • monitoring and guardrails
These are estimates/scenarios (not an invoice). Actual cost depends on context length, number of users, limits and retention policies.

Deployment / data / enterprise

Deployment channels

  • Mistral API
  • Le Chat (SaaS)
  • Self-host selected open-weight models (depending on variant)

Data policy

Training on data
Depends on service (API/SaaS) — check terms.
Retention
Depends on plan/service.
Data residency
EU/regions depending on hosting.
With self-host, data policy is on your side.

Enterprise readiness

Admin
SaaS/Enterprise depending on plan.
SSO/SCIM
Depends on enterprise plan.
Audit
Depends on enterprise plan.
DPA
Depends on agreement.
Certifications
Depends on agreement.
Good compromise: EU + ability to go self-host when scale grows.

Best use cases

  • self-host and on-prem (Mistral 3, Pixtral, Mixtral) with full data control and Apache 2.0 license
  • EU projects with emphasis on GPAI compliance and model documentation
  • document/image analysis with self-host (Pixtral 12B); code and analytics (Mistral 3, Codestral).

Strengths

  • Multiple models on Apache 2.0; good integration (vLLM, TensorRT-LLM) and inference optimizations.
  • Medium 3.5 / Small 4 — current commercial cloud variants; Mistral supports GPAI Code (EU).
  • Pixtral 12B — multimodal (text+image) in open; MoE in Large 3 — efficient inference.

Weaknesses / risks

  • Variant diversity requires task-specific selection; safety/guardrails are the deployer's responsibility with self-host.
  • Older models (7B, Mixtral) are previous generations; cloud variants — vendor lock-in and cost.

Current models (examples)

  • Mistral Medium 3.5, Mistral Small 4 (Mar 2026) — commercial cloud.
  • Mistral 3 (MoE/dense family), Pixtral 12B (VLM); Mistral 7B, Mixtral 8x7B — Apache 2.0.

Alternatives (if this model doesn't fit)

LL
Llama
larger open-weight ecosystem and hosting providers
OP
OpenAI
enterprise tooling and integrations
QW
Qwen
cheap open alternative (various scales)