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 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.
AI Model2026
Provides ComfyUI-ready INT8 MiniMax‑H3 checkpoints (conditioning encoder plus optional generation tail) for a Heretic-edited Qwen3‑VL‑32B source; preserves the vision tower in BF16 and uses row-wise ConvRot INT8 quantization to reduce VRAM needs for ~32GB GPUs. Not a full Transformers generation repository.
AI Model2026
ComfyUI-ready H3 conditioning encoder builds for Qwen3-VL-32B: a BF16 full-precision checkpoint, an INT8 ConvRot quantized checkpoint, and an optional generation tail (layers 50–63). Retains vision tower in BF16 and targets H3 workflows and lower-VRAM systems.
AI Model2026
An uncensored NVFP4-quantized text encoder for MiniMax-H3 video generation that fits on a single 16 GB GPU. Mixed-precision bake (mostly NVFP4, embedding left as INT8), preserves ConvRot rotation semantics, and includes the unrotate step required to avoid corrupted conditioning.
AI Audio2026
Generates complete songs (up to five minutes) from lyrics and a music description, producing 32 kHz stereo WAV with expressive vocals and long-range musical structure. Uses hierarchical LLMs fused with flow-matching/Flow-VAE synthesis for coherent arrangement and timbre; requires CUDA and integrates with Diffusers and SGLang-Omni.
AI Model2026
A customizable 30B-parameter Mixture-of-Experts LLM (3B active) in BF16 for low-latency, high-throughput agent workflows; supports speculative decoding (MTP/DSpark/DFlash) and up to 1M-token contexts. Released with open weights and recipes under OpenMDW-1.1, intended for post-training, domain adaptation, and research on NVIDIA GPU stacks.
AI Model2026
Open-weight 30B-parameter Mixture-of-Experts LLM with 3B active params, NVFP4-quantized checkpoint, and speculative-decoding support for long-context (up to 1M tokens) agentic, chat, reasoning and tool-calling workloads optimized for NVIDIA GPUs.
Post-training distribution-level objective that augments static Fréchet-distance losses with an adversarially learned representation and a real-feature whitening step to stabilize min–max optimization and avoid trivial feature amplification; targets one-step image generator post-training.
AI Model2026
Performs unified parsing of digital and camera-captured documents (layout, text, tables, formulas) using a ~1.2B-parameter vision–language model. Key differences: geometry-aware modeling, curvature-guided sampling, and content-structure decoupled training to handle real-world deformations without separate dewarping.