AI Model2026
A distilled 26M-parameter encoder–decoder LLM for on-device function-calling and tool use. Uses a pure-attention Simple Attention Network, provides open weights and local finetuning, and targets high-throughput inference on the Cactus runtime.
AI Model2026
Pretrained image-model checkpoint hosted on Hugging Face by Facebook (Meta) for vision experiments and transfer learning. Includes downloadable weights and metadata under CC BY‑NC 4.0 — suitable for research and prototyping but restricted for commercial use.
AI Dataset2026
A 228,557-example dataset of reasoning traces segmented into blocks with iterative, compressed "memento" summaries so LLMs can learn to manage long context. Includes a training-ready subset and a `full` subset with sentence/block-level annotations for research and SFT.
AI Dataset2026
Large-scale mid-training corpora for multimodal models: 10,809 ~60s video shards, caption splits (30s/60s/180s/>10min), 84 spatial-reasoning shards, and CSV mappings to source YouTube IDs. Small Parquet preview configs are provided for schema inspection.
AI Dataset2026
Provides 1,000,000 model-generated chain-of-thought traces and instruction–response pairs for fine-tuning and distilled supervision. Focused splits (coding, PHD-Science, General-Math, MultilingualSTEM), ~5B tokens, Apache-2.0 license.
AI Model2026
Runs the Bonsai family of quantized LLMs locally (including vision-capable 27B): provides scripts and demo UIs to run 1-bit and ternary Bonsai models on macOS (Metal), Linux/Windows (CUDA/Vulkan/ROCm), or CPU, with long context, tool-calling and an optional Open WebUI agent demo.
AI Dataset2026
Provides a unified 615k-hour English speech corpus for TTS training, aggregating 11 public datasets and web-sourced recordings into 16 kHz Opus WebDataset shards. Includes a quality-filtered core subset (510.1k hours), metadata splits, and mixed licenses across sources.
AI Dataset2026
Provides multi-turn agent trajectories with real tool executions and explicit <think> reasoning blocks for training and evaluating tool-calling agents. Contains two model-sourced configs (Kimi-K2.5, GLM-5.1) totaling ~14.7K samples — useful for SFT, agent-skill research, and tool-integration experiments.
AI Model2026
Converts text to natural-sounding speech across 600+ languages in a zero-shot way, with short-reference voice cloning and fine-grained voice-design controls; uses a diffusion language-model-style architecture to balance quality and very low inference latency.
AI Model2026
A dense 128B multimodal model with a 256k context window, configurable reasoning effort, and native function-calling for agentic workflows. Supports text+image input, multilingual output, and is released on Hugging Face under a Modified MIT license with revenue-based exceptions.
AI Model2026
Generates 48kHz multilingual speech from text using a tokenizer-free diffusion-autoregressive TTS architecture, supporting natural-language voice design, controllable cloning, and low-latency streaming. Notable for a 2B-parameter backbone and built-in AudioVAE super-resolution (16k→48k).
AI Model2026
Generates and iterates on long‑horizon agentic plans and code — designed to stay productive across many rounds of tool calls and experiments. Emphasizes iterative reasoning, stronger repo/terminal automation and code generation than GLM‑5, and can be served locally for research and autonomous-agent workloads.