Skip to content
Technology

Shieldstral

Shieldstral — Technology. definition: Multimodal safety classifier from Mistral AI (August 4, 2026), 3,8 Md actual parameters, open weights Apache 2.0, based on Ministral 3 + Pixtral vision encoder, 12 languages, 16 Go of VRAM in BF16; the moderation policy is supplied in natural language at inference time, not learned in the weights · limits: Binary yes/no output with no reasoning trace (auditability); mention/endorsement confusion (false positive on Voltaire); acknowledged weaknesses on obfuscated inputs, long documents, Arabic and Indonesian; as of August 6, 2026, neither a billed endpoint on La Plateforme nor an official Ollama

Type
Technology
definition
Multimodal safety classifier from Mistral AI (August 4, 2026), 3,8 Md actual parameters, open weights Apache 2.0, based on Ministral 3 + Pixtral vision encoder, 12 languages, 16 Go of VRAM in BF16; the moderation policy is supplied in natural language at inference time, not learned in the weights
limits
Binary yes/no output with no reasoning trace (auditability); mention/endorsement confusion (false positive on Voltaire); acknowledged weaknesses on obfuscated inputs, long documents, Arabic and Indonesian; as of August 6, 2026, neither a billed endpoint on La Plateforme nor an official Ollama
relations
12
Cited in
1 fiches

Neighborhood

Mistral AI Ministral 3 Llama Guard 4 Pixtral Qwen3Guard

← publishes

Mistral AI ORGANISATION high confidence stable Source ↗

→ is based on

Ministral 3 TECHNOLOGIE high confidence stable Source ↗

→ opposes

Llama Guard 4 TECHNOLOGIE high confidence timeless Source ↗

→ uses

Pixtral TECHNOLOGIE high confidence stable Source ↗

→ outperforms

Qwen3Guard TECHNOLOGIE high confidence evolving Source ↗

Cited in (1)