The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestMistral AIShieldstral
Mistral AIVision-Language

Mistral debuts Shieldstral, a 3B safety model

The open-weights multimodal moderation model brings content safety checks to both text and images under an Apache 2.0 license.

Aug 4, 2026
NotableApache 2.0

Mistral AI has released Shieldstral, a 3-billion-parameter open-weights model built specifically for content moderation across both text and images. According to Mistral's announcement, the model is aimed at helping developers screen inputs and outputs in AI applications, and it ships under a permissive Apache 2.0 license.

Unlike a general-purpose chat model, Shieldstral is a dedicated safety classifier. Its multimodal design means it can evaluate not just text but also images, an increasingly important capability as more applications accept mixed inputs and generate visual content.

Why it matters

Safety tooling has often been the least open part of the modern AI stack, with many moderation systems locked behind proprietary APIs. A compact, openly licensed model changes that calculus for teams that need to run checks themselves.

  • At roughly 3B parameters, it is small enough to deploy affordably alongside larger systems.
  • The Apache 2.0 license permits commercial use and modification.
  • Multimodal coverage lets it handle both text and image moderation in one model.

The release fits a broader industry pattern of shipping guardrail models as companions to frontier systems, echoing efforts like Meta's Llama Guard. For developers building on open weights, having a moderation layer from the same ecosystem lowers the friction of assembling a safer pipeline without giving up control of their data.

Sources

    OlderinclusionAI ships Ling-3.0-flash, an MIT-licensed MoE modelinclusionAI · Text / LLM · 4 months agoNewerNVIDIA's Nemotron 3.5 Lightning trims MoE for speedNVIDIA · Text / LLM · 4 months ago

    Get the model


    Specs

    Parameters3B
    LicenseAPACHE-2.0

    Can you run it?

    Runs on just about anything.

    • BF16 (as published)

      8 GB GPU (RTX 4060) · 16 GB laptop or Mac

      7.7 GB
    • 8-bit

      8 GB GPU (RTX 4060) · 16 GB laptop or Mac

      4.9 GB
    • 4-bit

      Any modern laptop · 8 GB RAM

      3.5 GB
    Your machine
    • Precision not published; assuming 16-bit (BF16) weights.
    • Parameter count is an editorial estimate — no weights metadata on Hugging Face.
    • KV cache approximated from model size (no config metadata) at 8,192 tokens.

    Estimated from the parameter count. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


    Modalities

    Any-to-AnyVision-Language

    The Weekly Weights

    Every open release that mattered, one email a week.

    0 comments

    No comments yet. Be the first to weigh in.

    More from Mistral AI

    All Mistral AI releases →
    Mistral Large 4 (1T-A52B)
    Mistral AI/Text / LLMDatacenter

    Mistral Large 4 Arrives as a 1T-Parameter MoE

    The new flagship pairs a trillion total parameters with 52B active per token and adds vision alongside multilingual text.

    Oct 29, 2026
    Timestral-209M
    Mistral AI/Text / LLMRuns anywhere

    Mistral's Timestral brings forecasting to open weights

    The French lab releases a compact 209M-parameter foundation model aimed squarely at time-series prediction.

    Oct 14, 2026
    Mistral AI/Text / LLMDatacenter

    Mistral Large 4 arrives as a trillion-param MoE

    The new flagship is a sparse mixture-of-experts model with roughly 49B active parameters per token.

    Oct 6, 2026

    More in Vision-Language

    All Vision-Language →
    SenseNova-Loop-9B
    SenseTime/Vision-LanguageRuns on a laptop

    SenseTime Releases Looped 9B Multimodal Model

    The company's new SenseNova-Loop-9B pairs vision-language capabilities with a recurrent looped-transformer design.

    Nov 17, 2026
    Hunyuan Hy4-flash
    Tencent/Vision-LanguageMulti-GPU server

    Tencent releases Hunyuan Hy4-flash multimodal MoE

    The lightweight, mixture-of-experts model adds vision and text instruction-following to Tencent's open Hunyuan lineup.

    Nov 17, 2026
    Mirage
    Cactus Compute/Vision-Language

    Cactus Compute's Mirage brings vision tasks on-device

    The new open vision-language model targets detection, segmentation, and visual Q&A without relying on the cloud.

    Nov 5, 2026