ISTA-DASLab prefiller model boosts Qwen3.8-27B low-quant accuracy

Ornith released Ornith-1.5-122B, extending Ornith-1.0, which was built on Qwen3.5 and Gemma4 with additional continued pretraining and mid-training. The lab frames it as a step toward foundation models built through end-to-end self-improvement.

Spotify says it removed over 50 million AI-generated "slop and spam" tracks from user recommendations in the past year, following 75 million spam takedowns the year before. Artists flagged with an "AI Persona" badge are blocked from algorithmic and editorial playlists.

A community LoRA for MOSS-Image-SFT cuts generation to 3 steps, hosted on Hugging Face under akhaliq/MOSS-Image-SFT-4-steps. The repo labels it a "3-step draft tier" — billed as the fastest MOSS-Image variant that stays clean.

Cohere's CALIBER method trains a reasoning model to estimate its chance of success before reasoning and assess its answer afterward; on its math benchmark, a calibrated 7B model cuts calibration error by more than half versus the best single-confidence baseline.

Google Research added a Matrix Execution Unit to its 32-bit open-source Coral NPU RISC-V architecture, delivering up to 1 TOPS while targeting 10 mW average power for edge inference. The release also ships the v1.0 open-source IREE/MLIR toolchain, and Synaptics is unveiling chips for wearables and industrial devices built on Coral.