Gemma 4 31B Abliterated — quality-preserving guardrail removal for Google's most capable open model. Apache 2.0. Runs on Apple Silicon via MLX.

Personalize Gemma 4 and make it real-time on Apple Silicon. Fine-tune on your conversations, serve at 111 tok/s. No cloud required.

Benchmark Gemma 4 E2B on Apple Silicon: MLX (mlx-lm) vs GGUF (llama-server), with TTFT, tokens/sec, and memory.

Drop-in KV cache compression for MLX on Apple Silicon. Brings PolarQuant (Google, ICLR 2026) to mlx-lm with first-class Gemma 4 support: MatFormer, dual head_dim, hybrid sliding/global attention, cross-layer KV sharing. 3-bit → 4.8× smaller cache, 0.995 logit cosine @ 4-bit.

M.A.R.A. is a Multi-Agent Research Analyst. A self-correcting multi-agent research system built with LangGraph, Gemma 4, and Streamlit.

Hands-on benchmark of Google Gemma 4 in LM Studio, with prompts, Gradio evaluator, saved outputs, and a published companion report.

Agentic Coding MCP Server: Local Gemma 4 powered coding assistant integrated with Zed editor

Nairobi-first edge AI simulation for urban flood resilience. Orchestrating multimodal agents (Gemini 2.5/Gemma 4) to navigate the Mbagathi Basin during infrastructure failure. Offline-first, spatial-aware, and mission-critical.