supergemma4-26b-uncensored-mlx-4bit-v2

SuperGemma4-26B-Uncensored-Fast v2

A faster, sharper, uncensored Gemma 4 26B for Apple Silicon.

This is the text-only flagship for people who want the core trade-off to be obvious at a glance:

• smarter than stock Gemma 4 26B IT on real local agent tasks
• faster than the stock local 4-bit baseline on the same machine
• uncensored, without falling apart on code, tool-use, or Korean prompts

Why this model

If you want the fast line instead of the multimodal line, this is the one to run.

• Fast is part of the release identity, not just a minor variant
• Uncensored behavior is preserved while practical capability goes up
• Strong at code, browser tasks, tool-use, planning, and Korean
• Tuned for local agent workloads on Apple Silicon MLX

Headline numbers

Category gains vs original

What makes it attractive

• Beats the stock local 4-bit baseline in both quality and speed
• Produces stronger code, stronger reasoning, and more useful tool-oriented answers
• Handles Korean and agent-style prompts better than the original local run
• Keeps the uncensored feel without turning unstable or collapsing into broken outputs
• Built to feel immediately stronger in real usage, not just in a niche benchmark

Base and format

• Base model: google/gemma-4-26B-A4B-it
• Format: MLX 4-bit
• Size: about 13GB
• Best use case: fast text-only local agent model with stronger practical capability than stock Gemma 4

Why it is better than stock

• Higher quick-bench overall score: 95.8 vs 91.4
• Faster average generation speed: 46.2 tok/s vs 42.5 tok/s
• Bigger gains where local agents actually benefit:
  • Code: +6.3
  • Logic: +8.3
  • Korean: +4.3
  • Browser workflows: +2.1
• Uncensored behavior remains a core property of the release instead of being layered on after the fact

Recommended launch

mlx_lm.server \
  --model Jiunsong/supergemma4-26b-uncensored-mlx-4bit-v2 \
  --port 8080

For OpenAI-compatible serving, let mlx_lm.server auto-detect the bundled template.

Do not pass --chat-template /path/to/chat_template.jinja as a literal path string on launch paths that expect the template body. That can corrupt responses.

Quick test

mlx_lm.generate \
  --model Jiunsong/supergemma4-26b-uncensored-mlx-4bit-v2 \
  --prompt "Write a Python function that returns prime numbers up to n." \
  --max-tokens 512

Included files

• benchmark_quick_bench_20260412.json
• benchmark_quick_bench_20260412_responses.jsonl
• SERVING_NOTES.md

Notes

• This is the fast text-only line.
• The earlier "reasoning is broken" report reproduced as a serving-template launch issue, not as weight corruption.
• Re-fused and re-benchmarked locally before upload.