Mistral 7B Instruct v0.2

Item: Mistral 7B Instruct v0.2
Rating: 2
Author: Ali Madjaji

Mistral AI · France

Based on published licence terms, Mistral 7B Instruct v0.2 is an EU-origin open-weight model under standard Apache 2.0 — commercial deployment and self-hosting are permitted without field-of-use restrictions. Training-data opacity is the primary EU AI Act Art. 53 gap, but the French controller, absence of CLOUD Act exposure, and Mistral's published DPA make this a strong baseline for regulated EU workloads.

Licence facts

Parameters: ~7.3B dense
Architecture: Decoder-only transformer with Grouped-Query Attention and RoPE (theta 1e6); full attention (sliding-window removed vs v0.1)
Context length: 32K tokens
Tensor type: BF16
Chat template: [INST] / [/INST]
Released: 2023-12

Known risks

Training-data transparency gap: Mistral's help centre explicitly states training datasets are not disclosed; the model card and paper describe only 'open web' sources with no corpus breakdown, filtering description, or opt-out mechanism — the primary Art. 53(1)(d) compliance gap, inherited by downstream deployers.
No built-in safety guardrails: the model card notes the absence of moderation; consumer-facing EU deployments must add their own filtering and downstream-provider documentation under the AI Act.
Fine-tune provenance is undisclosed: the instruct-tuning dataset is not published, which may matter for sectoral obligations (e.g. DSA Art. 34 risk assessments) or for copyright-TDM opt-out audits.

Sources

→ Model Card → Mistral 7B Paper (arXiv)→ Mistral AI — training-data disclosure note → Stanford FMTI — Mistral 7B transparency report → Mistral AI Data Processing Addendum