Cactus Compute's Mirage Squeezes Six Vision Tasks Into a 22 MB Phone Model
Cactus Compute's Mirage packs scene description, detection, OCR, segmentation and structured extraction into a 22 MB model that runs on-device in 74 ms.
20 stories
Cactus Compute's Mirage packs scene description, detection, OCR, segmentation and structured extraction into a 22 MB model that runs on-device in 74 ms.
Anymove extracts 3D motion capture from a single AI-generated video using a dual-camera trick and a DINOv2 joint reader, outputting rigged GLB animations for about $3 per move.
Tencent's new Hunyuan Hy4 takes the top spot on CWE-Bench-AA at 76% and sits on the capability-vs-cost Pareto frontier despite being trained with offensive security data stripped out.
A new benchmark and framework called FF-TRUST tackles sleep staging when labels are noisy and training data comes from mismatched clinical sources.
Sakana AI shifts from merging open models to training its own foundation model from scratch, with Sawara Small pretraining starting next month.
Tencent Hunyuan open-sourced Hy Memory, a dual-system cognitive memory engine with three specialized pipelines for chat, productivity, and coding agents.
A new graph attention based reconstruction model sets accuracy and latency records on three industrial defect benchmarks using only one to eight normal samples.
Google researchers give harness optimizers a structured log of past edits, and that history beats starting from the evolved harness alone when the model or benchmark changes.
TUM's DeepBA turns bundle adjustment into a 2M-parameter graph network. It recovers cameras, focal lengths and depth from 2D matches alone, starting from scratch, and gets close to billion-parameter models.
AnisoFlow makes the prediction target of flow models fully learnable, cutting ImageNet FID from 1.90 to 1.55 and killing patch seams in 8-bit inference.
miHoYo's Gamer model turns a Qwen3-VL backbone into an agent that plays Honkai: Star Rail with keyboard and mouse at human speed.
Math Inc released an executable Lean model of Linux's WireGuard handshake, a first step toward AI-generated proofs that kernel C code is correct.
Google.org is backing 148 doctoral researchers across 38 countries with $13 million, compute access, and direct mentorship from Google researchers.
vLLM retuned DeepSeek-V4.1-Flash for NVIDIA's Vera Rubin NVL72 rack, hitting 1.7M tokens/s per GPU and 58x GB300 throughput at interactive latencies.
Cognition rebuilds Devin's Teams tier with free SWE-2 sessions, SSO, per-user spend caps, and a HIPAA-ready BAA available without an enterprise contract.
MiniMax released a 33B parameter open-weights image model that generates and edits 16 megapixel images in just 6 sampling steps.
Cerebras posted record Q3 revenue with cloud growing 265%, landed two new hyperscalers including a top three, and locked up 700 megawatts of capacity.
Thinking Machines just dropped the Inkling 2 Small weights on Hugging Face under Apache 2.0, opening up its 594B agent-first MoE for self-hosting and fine-tuning.
Anthropic puts Claude inside Google Meet calls as a visible participant that drafts docs, timelines, and action items while you talk.
Kaggle launches a human-voted leaderboard where frontier models go head-to-head on charts, floorplans, websites, videos, and ASCII art.