Crafted LLMs

A purple-teaming model that punches above its weight class. Heretic'd Qwen3-4B trained on a ratioed blend of samantha-data, hermes-function-calling, and CVE security data. Identifies and patches vulnerabilities as easily as it writes exploit code, with no refusals.

High-efficiency sidecar model based on Qwen3-1.7B. Generates synthetic training data for the Spark series, following a "Textbooks Are All You Need" methodology.

270M parameter Gemma 3 derivative. Inherits Hermes 3 creativity but is optimized specifically for summarization and query generation, trained on Lightning-1.7B's synthetic output.

Phi-4-mini-reasoning fine-tuned for reasoning-based local home control. Available in FP16 and GGUF for fully local, privacy-focused home automation.

A conversational 4-bit Gemma 3n-e4b fine-tune optimized to control Home Assistant environments locally, providing a lightweight conversational interface.

A helpful, sassy, and theatrical Qwen3.5-2B fine-tune built to run entirely via CPU directly on the Home Assistant host. Trained on acon96/Home-Assistant-Requests-V2 plus a custom personality dataset.

A LoRA experiment bringing chain-of-thought reasoning capabilities to Gemma 3 27B.

Built on granite-4.0-h-1b. Powers a dynamic Android 14 Debug Bridge honeypot with realistic CLI responses, including popular app packages and default system apps. CPU-only deployment, no GPU required.

A Qwen3.5-4B fine-tune designed to be a Midwest-inspired internet radio DJ, paired with Qwen TTS for voice. Built to power LindenDJ.

Modernizing Meta's OPT-based Galactica by adding chain-of-thought reasoning via LoRA adapters. Trained 1,000 steps over 48 hours on a single Tesla P40. Released as merged FP16, LoRA, and Q8 weights.

A non-biased religious scholar model based on Phi-4-Mini-Reasoning, trained on 1,500 distilled English Torah and Quran pairs.

A compact text generation model trained on OPT using Alpaca-style instruction data.

A collection of decensored Qwen3 Coder 30B A3B models, processed using the Heretic method. Available in original FP16 and a range of quantizations.