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

Key Concepts

• Super Gemma 4 26B Uncensored: A community-driven, fine-tuned version of Google’s Gemma 4 26B model, optimized for fewer restrictions and improved performance in agentic workflows.
• Mixture of Experts (MoE): An architecture where only a subset of parameters (approx. 3.8B out of 25B) are active during inference, balancing power with efficiency.
• MLX (Machine Learning Exploration): An Apple-native framework designed for efficient machine learning on Apple Silicon.
• GGUF (GPT-Generated Unified Format): A file format for storing models, optimized for inference on various hardware (Windows/Linux/Mac) using llama.cpp.
• Agentic Workflows: Tasks involving tool use, code generation, browser automation, and logical reasoning, rather than simple conversational chat.

---

1. Main Topics and Technical Details

The video introduces Super Gemma 4 26B Uncensored (MLX 4-bit V2), created by Jun Song. While based on Google’s official instruction-tuned Gemma 4, this version is modified to be "uncensored," removing restrictive guardrails that often hinder local agent tasks.

• Architecture: The model utilizes a Mixture of Experts (MoE) design, maintaining 25B total parameters with only 3.8B active during inference.
• Performance Claims: According to the creator’s benchmarks, the model achieves a "Quick Bench" score of 95.8 (vs. 91.4 for the baseline) and an average generation speed of 46.2 tokens per second (vs. 42.5).
• Scope: This is a text-only model. It is explicitly designed for coding, tool use, browser tasks, and planning, rather than multimodal (image) understanding.

2. Implementation and Frameworks

The video outlines two primary ways to deploy the model:

A. Apple Silicon (MLX)

• Installation: Use pip install -U MLX-LM.
• Execution: Run the server using MLX_LM.server with the model ID and --port 808.
• Crucial Note: The author warns against manually forcing a chat template path, as the model relies on an auto-detected bundled template to prevent response corruption.

B. General Ecosystem (GGUF)

• For Windows, Linux, or users preferring llama.cpp tools (LM Studio, Jan, Open Web UI), a Q4_K_M GGUF version is available (approx. 16.8 GB). This version uses a "neutral embedded template" to prevent prompt drift into unintended coding or tool-calling modes.

3. Real-World Applications: Agent Integration

The model is designed to be plugged into "Agent Shells" that support OpenAI-compatible endpoints:

• Hermes Agent: A terminal-first agent framework. By pointing Hermes to the local MLX server, users gain a local, uncensored reasoning engine with native function calling and system prompt support.
• Open Claw: A personal assistant and task runner. Users can replace cloud-based APIs with the local Super Gemma endpoint to handle reasoning tasks, with the option to adjust memory limits via sysctl settings.

4. Key Arguments and Perspectives

• Practicality over "Edginess": The presenter argues that this fine-tune is significant because it avoids the common pitfall of uncensored models becoming "chaotic" or "role-play heavy." Instead, it maintains the structural integrity required for professional agent work.
• Native Agent Support: The presenter emphasizes that Gemma 4’s base architecture (native function calling and 256K context) makes it an ideal candidate for agentic fine-tuning, as it doesn't require "forcing" agent behavior onto a model not built for it.

5. Notable Quotes

• "The question was never really whether Gemma 4 had the right architecture. The real question was whether somebody would make a version that feels more open, more permissive, and a bit more useful for actual local style workflows."
• "This does not look like one of those weird uncensored fine-tunes that becomes chaotic... It is trying to be an uncensored model that is still actually useful."

6. Synthesis and Conclusion

Super Gemma 4 26B represents a bridge between high-performance base models and the specific needs of local power users. By stripping away restrictive filters while preserving the MoE architecture and native function-calling capabilities, it provides a robust foundation for local agent automation. Whether deployed via MLX on Apple Silicon or GGUF on other platforms, the model serves as a practical, high-speed alternative to cloud-based reasoning engines for developers and power users.