Tag

OpenShell

Runs autonomous AI agents in isolated containers with declarative YAML policies that enforce filesystem, network, process, and inference controls. Provides credential providers, inference routing, hot-reloadable network/policy updates, and formal verification of policy changes for auditable agent execution.

Rapid-MLX

Runs local AI models on Apple Silicon as an OpenAI‑compatible server, emphasizing low latency, prompt caching, and reliable tool-calling. Optimized for M1–M4 Macs with multimodal support and drop‑in compatibility for IDEs and agent frameworks.

ANE Training — Backpropagation on Apple Neural Engine

Runs transformer forward+backward training directly on Apple's Neural Engine by reverse-engineering private ANE APIs; includes per-layer ANE kernels, INT8 optimizations and benchmarks. Proof-of-concept only—relies on undocumented APIs, has low utilization and CPU fallbacks.

OBLITERATUS

Identifies and surgically removes the internal activation directions that trigger refusal behavior in large language models, with one-click options on a HuggingFace Space or a local Python API. Combines multiple extraction methods (SVD, whitened SVD, sparse autoencoders), reversible steering, and analysis-informed verification to quantify capability and refusal trade-offs.

whichllm

Detects your GPU/CPU/RAM and ranks HuggingFace models that will actually run and perform best on your machine, using live, recency-aware benchmarks and a one-line CLI for simulation, planning, and instant local runs.

Qwen3.5-9B-Claude-4.6-HighIQ-THINKING-HERETIC-UNCENSORED

Community fine-tuned multimodal Qwen3.5-9B using Claude 4.6 distilled data to change the model's 'thinking' behavior; offers an uncensored 'heretic' flavor with image-text-to-text I/O, benchmark comparisons, and deployment notes for inference frameworks.

Cosmos3-Nano (NVIDIA)

Generate text, images, video, audio and action/robot trajectories from combined text, image, video, audio and action inputs. A Mixture-of-Transformers omnimodal foundation model (Cosmos3‑Nano, 16B params) focused on Physical AI (robotics, AV, simulation) and optimized for NVIDIA GPU runtimes.

google/gemma-4-31B-it

Instruction-tuned Gemma 4 31B multimodal model that generates text from text+image inputs with up to 256K context. Dense 31B variant optimized for vision-language understanding, long-context reasoning, and coding; Apache‑2.0 licensed.

Needle

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.

turbovec

Compresses high-dimensional embeddings into low-bit TurboQuant indexes for fast, memory-efficient local vector search. Supports online ingest (no train/rebuild), SIMD kernels that match or beat FAISS, per-vector length-renormalization, and runtime allowlists — suited for privacy-sensitive, low-latency RAG.

Mistral Medium 3.5 128B

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.

Gemma 4 31B JANG_4M CRACK (v2) — dealignai

Multimodal image-text-to-text fork of Gemma 4 (31B) using a 'CRACK v2' abliteration — tuned for conversational vision inputs and thinking-mode support in JANG v2 safetensors format. Recommended to run in vMLX; published by dealignai.