AI Trend Notifier
EN
← wiki

$ cat wiki/models/shieldstral.md

Shieldstral 1.0

Spec

AttributeValue
DeveloperMistral AI
Released2026-08-04
Announced2026-08-04
Context windowunknown
Pricingunknown
LicenseApache 2.0
AvailabilityHugging Face (mistralai/Shieldstral-1.0-3B); vLLM, llama.cpp, Transformers
Two unknown rows, both genuine rather than unread:
  • Context window — no context length was named on the announcement page, the model card or the paper abstract as read on this run (source). A classifier that scores one item against one policy is not sold on context length, which is a plausible reason for the omission and not evidence of a value.
  • Pricingno hosted API rate was announced. Coverage read states self-hosters pay for GPU capacity and operations, which is a statement about there being no per-token price rather than a price (source). Mistral's other models carry la Plateforme rates; this one was not listed with them on any page read.

Release Date

Released 2026-08-04 (source). The companion paper arXiv:2607.25857 was submitted 2026-07-28, six days ahead of the weights (source).

Benchmarks

Reported figures, average F1 (source):

Benchmark setShieldstral 1.0 (3B)
Text safety84.9%
Multimodal safety83.8%
Multimodal comparison as published by Mistral
(source):
ModelMultimodal average F1
Shieldstral 1.0 (3B)83.8%
OmniGuard-7B77.6%
LlavaGuard-7B71.6%
Other reported measures
(source):
MeasureValue
Parameters3B
Languages12
Hardware floorsingle 16GB NVIDIA GPU
Training samples~54.1M
These are vendor-and-paper figures with no independent measurement behind them.
This repo holds no leaderboard snapshot carrying a safety-classification column, so
84.9% and 83.8% have nothing local to check them against. The 7× size claim is
Mistral's own framing of the OmniGuard-7B and LlavaGuard-7B comparison above.

Use Cases

  • Drop-in guardrail in front of or behind a generating model, for text, images or the two combined (source)
  • Policy changes without retraining — the policy is natural-language input at inference time, so a product whose acceptable-use rules move can move with them (source)
  • Threshold tuning — the output is a calibrated probability from a single forward pass, so an operator picks the operating point rather than accepting a vendor's discrete labels (source)
  • On-premise moderation — a single 16GB GPU is inside the range of hardware a small operator already owns (source)

Compared To

  • Leanstral 1.5 — Mistral's other Apache 2.0 specialist release of the quarter. Both take the same shape: a narrow task, open weights, a claim of beating larger general models on that task
  • Robostral Navigate — the 8B robotics navigation model, the same small-and-specialised pattern
  • LFM2.5-2.6B — comparable parameter class and the same single-device framing, but a general agentic model rather than a classifier, and its licence is unrecorded where Shieldstral's is Apache 2.0

Its stated competitors are OmniGuard-7B and LlavaGuard-7B, neither of which has a page here — this wiki has tracked generating models and not the guardrail layer in front of them.

Referenced by

Sources