Tag

Granite Embedding 97M Multilingual R2

Produces 384‑dim multilingual (and code) embeddings with up to 32,768 token context, optimized for low‑latency production retrieval. Compact 97M model with ONNX/OpenVINO and vLLM/GGUF deployment options for edge and high‑throughput use.

Qwen3.6-27B

A 27B multimodal causal language model with a vision encoder and native long-context support (262,144 tokens). Optimized for repository-level coding agents and multimodal understanding; includes preserved "thinking" traces, multi-token prediction (MTP), and deployment recipes for vLLM / SGLang / Transformers.

Qwen3.6-27B-FP8

FP8-quantized 27B multimodal Qwen3.6 model weights in Hugging Face Transformers format — supports image/text/video inputs, native 262k token context (extensible to ~1M), and is compatible with vLLM/SGLang/KTransformers for efficient local serving and research.

HealthBench Professional

Benchmark dataset for evaluating clinician-facing chat assistants: physician-authored conversations plus rubric items, use-case and difficulty labels, specialty metadata, and a built-in canary to reduce benchmark contamination. Hosted on Hugging Face under an MIT license.

SynData

Provides a large-scale multimodal embodied dataset (vision, depth, hand/arm kinematics, tactile) captured with an exoskeleton glove and egocentric sensors; organized as clip-level Zarr volumes for manipulation, imitation learning, and vision–action research. Includes both high-precision glove measurements and natural bare-hand clips; sizable storage required.

DeepSeek-V4-Flash

A 284B-parameter Mixture-of-Experts LLM with only 13B activated parameters, designed for 1,000,000-token contexts. Uses hybrid compressed attention and mixed FP4/FP8 precision to reduce long-context KV-cache and per-token FLOPs; aimed at long-document QA, RAG pipelines, and local/high-capacity inference.

Qwen3.6-27B-Uncensored-HauhauCS-Aggressive

Provides a locally runnable, refusal-free variant of Qwen3.6-27B with multiple K_P GGUF quantizations and mmproj multimodal support. The Aggressive flavor skips preambles on edgy prompts—use when you want direct/raw responses for local research, red‑teaming, or offline workflows.

DeepSeek-V4-Pro

Generates conversational and reasoning outputs with support for million‑token contexts; uses a hybrid attention + MoE design to cut long‑context inference FLOPs and KV cache. Suited for long‑document retrieval, coding and complex reasoning; MIT licensed.

DeepSeek-V4-Flash-Base

A vision-oriented foundation checkpoint for low-latency inference — DeepSeek V4 base in safetensors with FP8 optimizations. Designed for fast image generation and embedding use in inference pipelines; verify license and FP8/runtime compatibility before production use.

H³D: High-quality Holistic 3D Editing Dataset

Provides instruction-based (before, after) structured 3D latents (SLAT) with aligned RGB views and natural-language edit prompts for training and evaluating instruction-following 3D editing models. Covers part-level semantic edits across seven edit types (deletion, addition, modification, scale, material, color, global) and supplies shard-based NPZ assets and loader code.

DeepSeek-V4-Pro-Base

Base image-generation foundation model tuned for visual search and prompt-guided synthesis, intended as a compact starting point for local inference or fine-tuning. Emphasizes easy integration into image pipelines and suitability for downstream adaptation.

LLaDA2.0-Uni

Unifies multimodal image understanding, text-to-image generation, and instruction-based editing in a single diffusion LLM using a Mixture-of-Experts backbone, SigLIP-VQ discrete tokenizer, and a distilled diffusion decoder enabling fast (8-step) decoding; full-generation needs ~47GB GPU RAM.