Provides 104.9M curated image–text pairs with precomputed embeddings, structured annotations and pre-encoded VAE latents for text-to-image pretraining and retrieval. Combines filtered web sources and synthetic samples with multi-model re-captioning, deduplication and safety filters; Apache-2.0.
Converts latent representations into high-resolution images by using a conditional pixel-space diffusion decoder that merges decoding and upsampling into a single generative step. Released checkpoints are 4-step distilled (2k and 2kto4k variants) and pair with specific VAE/encoder weights; license restricts use to non-commercial research.
A 40B GGUF-quantized Qwen3.6 variant fine-tuned with Claude 4.6 Opus and Deckard/Heretic datasets for multimodal image-text-to-text tasks. Offers 256K context, custom NEO-CODE Di-IMatrix quants for long conversations and coding, optimized for local inference and creative/coding use cases; safety alignment removed.
Transforms pretrained latent-diffusion priors into pixel-space diffusion models by removing the VAE and training shallow pixel layers on LDM-generated synthetic images — enabling fast convergence, native 4K output, and low-data training on 8 GPUs.
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.
Provides tick-aligned Counter-Strike 2 player POV video clips with per-tick inputs and world-state sidecars — near-lossless 1280×720@32fps video, per-player stereo audio, and parquet indexes for event/kill/round filtering; suited for RL, video classification and clip mining.
A 30B mixture-of-experts multilingual translation model supporting 33 languages and instruction-following translation. Offers MoE architecture, fast-thinking mode, and quantized/deployment-ready variants for production translation and subtitle tasks.
Generates high-quality Japanese speech from text with zero-shot voice cloning and emoji-based style controls; uses a flow-matching diffusion transformer over DACVAE continuous latents, includes a duration predictor and integrated SilentCipher watermarking. Japanese-only.
Research-focused text-to-image foundation model that prioritizes training efficiency: a 3.8B-parameter architecture trained on an 800M image-text corpus with mixed-resolution learning, FLUX.2 VAE, RL tuning, and a distilled 4-step Lens-Turbo for fast high-resolution generation.
Delivers image and video generation, editing, and understanding inside a single 3B-parameter multimodal model trained from scratch with a multi-task recipe. Notable for strong unified benchmarks at 3B scale; inference requires large GPU memory (≈40GB+ VRAM).
A 4-step distilled variant of Microsoft's Lens foundational text-to-image model for fast, high-resolution image synthesis. Optimized for mixed-resolution inference up to 1440×1440, GPT-OSS text features and FLUX.2 latents, intended for low-latency prototyping and research under an MIT license.
Multilingual streaming ASR that transcribes 40 language-locales using a cache-aware FastConformer‑RNNT architecture. Supports language-ID prompting (or auto-detect), punctuation/capitalization, and configurable chunk sizes to trade latency vs. accuracy for production transcription and streaming voice agents.