ModelsRadar
AISI publishes evaluation results with their experimental configurations
The UK AI Security Institute (AISI) shares results through Evaluation Cards, including the configurations needed to interpret differences between scores.
Helena Prado
Models & research
An editorial pseudonym for Shift’s AI-assisted desk. Editorial policy
17 articles
ModelsRadar
TII launches Falcon ASR with Arabic and Emirati evaluations
Technology Innovation Institute's speech recognition model supports Portuguese, but its published accuracy results cover Arabic, Emirati speech and English.
ModelsRadar
Open TTS Leaderboard measures speech accuracy and playback delay
The text-to-speech comparison separates intelligibility, reference-voice similarity and generation speed; these measurements do not directly measure naturalness or listener preference.
ModelsRadar
Beam activates selected experts but keeps all weights in memory
Reflection's text model activates 23 billion of its 501 billion parameters per token; deployment also needs memory for stored context and concurrent requests.
ModelsRadar
Swiss AI publishes Apertus 1.5 evaluation settings alongside multimodal results
The text-generating models accept images and audio; Every Eval Ever records document generation settings, model access and metric interpretation.
ModelsRadar
Muse Spark Open samples video frames before processing input
Meta's model samples video at two frames per second by default, capped at 96 frames; denser sampling increases memory use and input-processing time.
ModelsRadar
VocabX compresses grading output while keeping scores separate
Dharma-AI adapts an essay grader's vocabulary around generation cost, treating repeated JSON data structures separately from the grades they contain.
ModelsRadar
Thinking Machines Lab launches Inkling 2 Small on Tinker with closed weights
The model is available for inference and fine-tuning; the planned Apache 2.0 weight release depends on external safety testing.
ModelsRadar
Liquid AI launches six LFM3 models with different input and output modalities
Text, vision, audio, Omni, and decision variants share a backbone; image, video, and speech support depends on the selected model.
ModelsRadar
Mistral Large 4 adds text-and-image inference with three reasoning levels
Transformers supports low, medium, and high reasoning effort; none still selects high. The weight license restricts model-as-a-service offerings.
ModelsRadar
Project ProtoStar launches open challenges to test scientific data reuse
The initiative seeks to make scientific datasets, models, and code reusable, starting with urban climate forecasts compared against historical averages and linear trends.
ModelsRadar
Ai2 introduces ToyBlocks to study how robot skills transfer
A simulated arm and tabletop blocks let researchers vary training demonstrations and evaluate performance on tasks held out of training.
ModelsRadar
Open ASR Leaderboard adds Egyptian and Nigerian English evaluations
The automatic speech recognition benchmark adds 1.34 hours of Egyptian-accented English and 1.55 hours of Nigerian-accented English, both conversational and spontaneous.
ModelsRadar
NVIDIA adds four Cosmos checkpoints for controlled simulation video
The Cosmos 3 variants generate driving observations or action-conditioned video, including synchronized camera observations and LiDAR outputs for 3D sensing.
ModelsRadar
IBM updates Granite Switch speech and decision components
Granite Switch 4.1.1 keeps the base model unchanged and updates speech recognition across 10 languages with punctuation and capitalization.
ModelsRadar
LeRobot adds dataset streaming and live policy instructions
Version 0.7.0 trains without a full dataset download; questions and autonomous goals require a text-generating policy or an external planner.
ModelsRadar
Transformers uses lower-precision weights to reduce memory
Quantization stores learned model parameters at lower precision; method compatibility and hardware support determine which options can run.