AI Audio2026
Local-first voice workflows for cloning, multi-engine TTS/ASR, video dubbing, dictation, transcription and audiobook production across 646 languages. Desktop app with a local OpenAI-compatible API, engine catalogue (TTS/ASR/LLM), and explicit opt-ins for remote features to keep audio and projects on-device.
AI Infra2026
Provides high-performance CUDA/CUTLASS kernels implementing Kimi Delta Attention (KDA), accelerating KDA prefill on SM90+ (Hopper) GPUs. Integrates as a drop-in backend for flash-linear-attention, supports native variable-length batching, and targets K=V=128; requires CUDA 12.9+/PyTorch 2.4+.
AI Deploy2026
Provides reusable Docker compose files, scripts and benchmarked configs to serve modern LLMs (Qwen, Gemma, etc.) on 1–2 NVIDIA RTX 3090/4090/5090 GPUs. Multi-engine (vLLM, llama.cpp, ik_llama), measured TPS/context tradeoffs, and validated single/dual‑GPU recipes.
AI Infra2026
Native local inference engine for DeepSeek V4 Flash (also supports GLM 5.2 and PRO on high‑memory machines). Focused features include model-specific loading, SSD expert streaming, asymmetric routed-expert 2-bit quant support, multi-GPU/tensor/pipeline parallelism, and an OpenAI-compatible server plus a native coding agent.
AI Infra2026
High-throughput LLM inference engine for agentic workloads, combining a local‑SPMD static compiler for parallelism, a C++ scheduler with a Python execution plane and type‑safe KV‑cache reuse, pluggable high-performance kernels (including an MLA implementation), and a low‑overhead AsyncLLM entrypoint for production GPU inference.
AI Train2026
End-to-end Python framework for training and serving NVIDIA's Cosmos world models (Cosmos3), integrating distributed training (FSDP/TP/CP/PP), DCP/safetensors checkpoints, dataset adapters, multiple inference backends, online serving, and agent skills.
AI Infra2026
Streams routed experts from disk and treats storage, RAM and VRAM as a single inference hierarchy so MoE models (hundreds of billions to trillions of parameters) can run on consumer or heterogeneous hardware; single-file C engines with zero runtime deps and optional Vulkan/CUDA/Metal backends, speed governed by placement and disk I/O.
AI Model2026
Runs a full 27B-class language model using end-to-end binary (1.125-bit) weights, cutting FP16 size to ~3.9 GB. Key features: 262k-token context, custom 1-bit kernels for Apple MLX and CUDA, and an optional DSpark drafter for faster decoding. Best when memory footprint matters; trades some FP16 accuracy for on-device feasibility.
AI Model2026
Runs a full 27B-class Qwen3.6-derived LLM in a ~7.2 GB ternary/2‑bit format for on-device or single‑GPU text generation, retaining ~95% of FP16 performance and supporting a 262K‑token context. Designed for laptop/GPU deployment; exceeds typical phone memory limits.
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.