AI Model2026
27B multimodal reasoning model built on Qwen3.5-27B that preserves the base model's native multi-token-prediction head, full vision tower, and a 1,048,576-token YaRN context window. Designed for agentic tool use, long-context reasoning, and research deployments; released under Apache-2.0.
AI Model2026
End-to-end 0.8B multimodal OCR and page-level document parser that converts page images into structured Markdown (text, LaTeX formulas, HTML tables, and image crops). Post-trained from Qwen3.5-0.8B using mixed real/synthetic data and SFT+RL+OPD; achieves 96.58 on OmniDocBench v1.6.
AI Model2026
Agentic coding and long-horizon text generation via a 118B-parameter Mixture-of-Experts LLM with a 1,048,576-token context window. Features 256 routed experts, native preserved-thinking (reasoning) control, speculative decoding draft models, and quantized checkpoints for lower-cost serving.
AI Model2026
Generates 2048-d multilingual text embeddings for retrieval and semantic search, suited for RAG and dense retrieval. Pruned and distilled from the Ministral-3 family into a ~1.14B BF16 model, supports long contexts (up to 32,768 tokens) and optimized for NVIDIA GPU inference.
AI Model2026
Timestamp-aware realtime video→text model that processes incoming frames continuously, answers questions mid-stream or emits silence when evidence is insufficient, and can revise earlier outputs as new frames arrive. Built for timestamped multimodal interaction with a 256K context and an 11B-parameter backbone.
AI Model2026
Accepts text, image and audio inputs and generates text outputs for conversational, instruction-following and multimodal tasks; a sparse-MoE autoregressive model (975B total, 41B active) with BF16/NVFP4 support and local-deploy recipes.
AI Video Papers2026
Enables efficient, generalist video understanding by combining an Inflated 3D Vision Transformer and adaptive frame-resolution streaming with a scalable video data synthesis pipeline; ships as a fully open 4B-parameter MLLM that improves general, long-form, and streaming benchmarks.
Expands a Transformer’s residual stream into many parallel streams and introduces xHC to scale Hyper-Connections beyond N=4. Combines temporal feature augmentation with sparse residual updates (update k=4 of N=16) and xHC-Flash memory optimizations to raise downstream scores while cutting effective compute and memory traffic.
A looped-Transformer LLM series using Mixture-of-Experts (20B with 2B active; 6B with 0.6B active) that trades extra pretraining compute for repeated looping. Shows superior compute-efficiency versus matched-compute vanilla baselines and attains gold-medal performance on 2025 IMO and IPhO after a post-training pipeline.
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
Open preview checkpoint of a sparse Mixture-of-Experts causal LLM with ~314B total params (~13B active per token) and native 256K context for long-context multilingual text generation. Ships with custom modeling code (trust_remote_code) and a research/non-commercial use license.