Embodied AI2026
Generates robot manipulation actions from visual observations and text instructions using a 1.5B vision-language-action model. Uses streaming context and visual-token compression to cut per-step compute, runs a unified policy across tasks, and is open-sourced on Hugging Face under Apache-2.0.
Embodied AI2026
Predicts eight future [x,y,yaw] waypoints for language-conditioned embodied person-following using fused DINOv3 and SigLIP visual features; trained with quality-driven, DAgger-style self-evolving data and optimized for on-device inference (~5+ FPS, ~180 ms).
AI Model2026
Generates and reasons about multimodal physical-world content—text, images, video and action trajectories—conditioned on text, images, video and robot/vehicle action inputs. An edge-sized (4B) Mixture‑of‑Transformers omni-model optimized for single‑GPU inference and Physical AI tasks (image→video, action prediction, robot policy).
Efficient 4B-scale image generation and editing model family that pairs a lightweight VAE tokenizer (Mage-VAE) with a native-resolution multimodal diffusion transformer, reducing tokenization cost by an order of magnitude and enabling few-step high-resolution generation and editing.
AI Model2026
Efficient 4B native-resolution diffusion foundation model for text-to-image generation and instruction-based image editing. Uses a lightweight Mage‑VAE tokenizer and a 4B NR‑MMDiT backbone to produce 512–2048 outputs with low memory and fast inference; ships in Base, RL-aligned and few-step Turbo variants.
AI Image2026
Performs instruction-based image editing from reference images using a 4B native-resolution diffusion transformer; the Turbo variant uses 4-step distillation for interactive latency (≈1.02 s per 1024² edit on A100) while supporting semantic, appearance, structure-aware and restoration edits.
AI Model2026
Compact 3B-scale agentic LLM for multi-step tool use and reasoning, using a Looped Transformer to increase capacity without adding parameters; built for local deployment with configurable "thinking" modes and benchmark gains vs larger open models.
A PyTorch-native training framework for agentic reinforcement learning research that keeps researcher-facing code compact and editable. Uses an asynchronous loop to train multimodal and mixture-of-experts policies while never training on tokens the agent didn't generate; matches Megatron-style stacks under a comparable protocol and ships recipes and containers on GitHub.
AI Model2026
Provides a 2‑bit quantized build of Qwen3.6‑35B‑A3B for local serving via an OpenAI‑compatible HTTP API. Key features: 12.3 GB on disk, eschamoe mixed 2/3‑bit expert quantization with int8 dense layers, runs on a single 16–24 GB NVIDIA GPU and ships with Escha SGLang and ZML runtimes.
AI Dataset2026
Generates articulated motion for arbitrary rigged 3D skeletons from text prompts with a single unified model and no per-skeleton retraining. Uses topology-aware diffusion-transformer architecture and a large UniML3D motion dataset to generalize across bipeds, quadrupeds, and other topologies; suitable for text-to-animation workflows and rapid prototyping.
AI Video Papers2026
Real-time streaming multimodal foundation model that uses a codec-native tokenizer (Mage-ViT) to encode motion- and residual-rich regions from video I/P frames, reducing visual token usage by over 75% and enabling up to ~3.5× wall-clock inference speedup after training on ~560M images and 100M video frames.
AI Audio2026
Multilingual neural text-to-speech model (0.6B params) with zero-shot voice cloning and a bundled 44.1 kHz codec. Preview release targets 11 recommended languages and aims to deliver near-SOTA quality in a compact checkpoint suited for voice cloning and multilingual TTS prototypes.