Mistral's Shieldstral: 3B open-weights model for multimodal moderation

5 hours ago

Shieldstral is a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size by framing content moderation as a policy-adaptive question-answering task.
It accepts plain-language policies at inference time, unifying text and image safety evaluation without retraining, and runs efficiently on a single 16GB GPU.
Released under Apache 2.0, it delivers calibrated safety scores from a single forward pass, allowing thresholding or ranking by confidence.
Trained on heterogeneous public datasets converted into a unified format with contrastive examples, it teaches policy discrimination that generalizes to novel policies.
The model combines multiple fine-tuned checkpoints via SLERP to balance calibration and adaptability, and was built on the Forge platform.
Future directions include improving multilingual coverage, longer-document robustness, and broader multimodal safety.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token and a 1M-token context window, developed in under nine weeks.
It achieves competitive scores against much larger models on long-horizon coding benchmarks, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1.
All evaluation trajectories are publicly released at trajectories.poolside.ai for transparency and reproducibility.

Robostral Navigate is an 8B model for autonomous robot navigation using only a single RGB camera.
It achieves 76.6% success on unseen R2R-CE benchmarks, outperforming multi-sensor approaches.
The model combines pointing-based navigation with reinforcement learning for continuous improvement.

NVIDIA introduces Nemotron 3 Ultra, their most capable model with 550 billion total parameters and 55 billion active parameters.
Employs Mixture-of-Experts Hybrid Mamba-Attention architecture, LatentMoE, and MTP layers for speculative decoding and faster inference.
Supports inference time reasoning budget control and is pretrained in NVFP4 format.

BTL-3 is a 27B open-weight agent model designed for agentic coding, structural tool use, repository work, failure recovery, and long multi-turn execution.
It achieves high scores on benchmarks: 95.1% HumanEval pass@1, 88.5% BFCL v4 AST, and 88.1% LiveCodeBench v6.
A Compact edition (8.39 GB) packages the full model into a single file with 92.2% behavior retention, enabling native local inference.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token and a 1M-token context window, developed in under nine weeks.
It achieves competitive scores against much larger models on long-horizon coding benchmarks, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1.
All evaluation trajectories are publicly released at trajectories.poolside.ai for transparency and reproducibility.

Robostral Navigate is an 8B model for autonomous robot navigation using only a single RGB camera.
It achieves 76.6% success on unseen R2R-CE benchmarks, outperforming multi-sensor approaches.
The model combines pointing-based navigation with reinforcement learning for continuous improvement.

NVIDIA introduces Nemotron 3 Ultra, their most capable model with 550 billion total parameters and 55 billion active parameters.
Employs Mixture-of-Experts Hybrid Mamba-Attention architecture, LatentMoE, and MTP layers for speculative decoding and faster inference.
Supports inference time reasoning budget control and is pretrained in NVFP4 format.

BTL-3 is a 27B open-weight agent model designed for agentic coding, structural tool use, repository work, failure recovery, and long multi-turn execution.
It achieves high scores on benchmarks: 95.1% HumanEval pass@1, 88.5% BFCL v4 AST, and 88.1% LiveCodeBench v6.
A Compact edition (8.39 GB) packages the full model into a single file with 92.2% behavior retention, enabling native local inference.

Hasty Briefsbeta