In this video, I'll be telling you about SuperGemma 4, a community fine-tuned uncensored version of Gemma 4 26B, and how you can run it locally with MLX on Apple Silicon or use the GGUF version on other platforms for agent workflows.

--
Key Takeaways:

🚀 SuperGemma 4 is a community uncensored fine-tune of Gemma 4 26B aimed at text, coding, planning, and tool-use.  
🍎 The MLX 4-bit v2 release is built for Apple Silicon and can be run as a local OpenAI-compatible server.  
🤖 SuperGemma 4 pairs especially well with Hermes Agent for terminal-first local AI workflows.  
🧠 It can also be connected to OpenClaw as a local reasoning model for assistant and automation tasks.  
⚡ The model card claims improved speed, better benchmark scores, and stronger performance for code, logic, Korean, and browser tasks.  
💾 Realistically, 24GB of unified memory is the minimum comfortable floor, while 32GB or more is the better choice.  
🪟 A GGUF v2 version is also available for llama.cpp, LM Studio, Jan, Open WebUI, and other non-Mac setups.  
✅ Overall, SuperGemma 4 looks like one of the more practical uncensored Gemma variants for local agent work.

AICodeKing

SuperGemma-4 is a community fine-tuned uncensored version of Google's Gemma 4 26B model, optimized for local agent workflows on Apple Silicon via MLX. The MLX 4-bit V2 release by Jun Song on Hugging Face claims a QuickBench score of 95.8 vs 91.4 for the base and 46.2 tokens/sec vs 42.5. It focuses on text, coding, tool use, and browser tasks rather than multimodal capabilities. Setup involves installing MLX-LM, running a local server, and connecting it to OpenAI-compatible agent tools like Hermes or OpenClaw. A GGUF Q4_K_M variant (~16.8 GB) is also available for Windows/Linux users via llama.cpp, LM Studio, or similar tools. A key serving tip: let MLX_LM.server auto-detect the chat template rather than forcing one manually to avoid corrupted responses.

SuperGemma-4 (26B) UNCENSORED + Hermes,OpenClaw,OpenCode: THIS IS SO CRAZY!!!