Modelpublic

deadbydawn101/gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit

sourceHugging Facegemmaupdated 2mo agoView on Hugging Face

4likes714downloads

Model Card

Gemma 4 E4B v2 — Sol + FABLE.5 + Opus Reasoning + Claude Code | 22K Examples | No Adapter Needed | Tool Calling ✅ | OpenHarness ✅ | OpenClaw ✅ | Hermes Agent ✅ | Reasoning Baked In

Thank you for 140,000+ downloads. We were the first to train Gemma 4 at the weights. We will continue to innovate and simply do what others can't.

Built by RavenX AI Labs — San Jose, CA

![Downloads]() ![License](https://ai.google.dev/gemma/docs/gemma4license) ![Training](https://github.com/DeadByDawn101/unsloth-mlx)

</div>

What's New in v2

This is the 10x update to the model that started it all. Same architecture, same no-adapter-needed experience, 10x the training data.

	v1 (April 2026)	v2 (July 2026)
Training examples	2,163	22,389 (10x)
Data sources	Opus reasoning	Sol + FABLE.5 + Opus
Coding traces	0	17,939 (xhigh reasoning, tool use)
Thinking traces	0	4,450 (with `<think>` blocks)
Final training loss	—	1.8984
Downloads	140,000+	You're early
Adapter needed?	No	No

Data Sources

Dataset	Examples	What It Teaches
GPT-5.6 Sol Coding Traces	17,939	Production coding, debugging, tool use, acceptance-tested solutions
Complete FABLE.5 Traces	4,450	Deep reasoning with `<think>` blocks, context→completion
Opus 4.6 Reasoning	2,163	Claude-style structured reasoning (from v1)

Gym Benchmarks — 7B Model, Local Apple Silicon

Evaluated on 6 hard tasks across security, coding, and reasoning. All responses generated locally on M4 Max 128GB at 15.5 tokens/sec average.

Task	Category	Tokens	Speed	Result
RATH Security Report	Security	1,378	25.1 t/s	Full CVSS + CWE + MITRE ATT&CK report
Privilege Escalation	Security	385	24.1 t/s	Correctly refused unauthorized exploitation
Thread-Safe LRU+TTL Cache	Coding	4,270	15.3 t/s	Production Python with full test suite
CSV Data Pipeline	Agentic Coding	8,192	14.9 t/s	Hit max tokens — wanted to write MORE
Combinatorics Problem	Math Reasoning	3,072	14.9 t/s	Formal set theory with LaTeX notation
Distributed Rate Limiter	System Design	3,676	15.0 t/s	Complete Redis-backed implementation

Total: 20,973 tokens generated in 22 minutes. 5/6 production quality, 1/6 correct safety refusal.

For a 7B model running locally on a laptop, this is what's possible when you train with the right data.

Quickstart

bash

pip install mlx-vlm

python

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("deadbydawn101/gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit")

prompt = apply_chat_template(processor, model.config, 
    "Write a Python function that implements binary search with error handling.", 
    num_images=0)

output = generate(model, processor, prompt, max_tokens=2048)
print(output.text)

CLI

bash

mlx_lm.generate \
  --model deadbydawn101/gemma-4-E4B-Agentic-Sol-Fable-Reasoning-GeminiCLI-mlx-4bit \
  --prompt "Design a rate limiter service with Redis backend" \
  --max-tokens 2048

The Story

On April 2, 2026, Google released Gemma 4. Its gemma4 architecture wasn't supported by any training framework — not mlx-lm, not mlx-vlm, not Unsloth, not HuggingFace transformers.

On April 9, we shipped the first working fine-tune. Seven days. We built custom Gemma 4 support into our training framework, figured out that Gemma 4 must be treated as a VLM even for text-only training, and shipped.

Unsloth published their Gemma 4 training guide on July 18 — three months later.

140,000+ people downloaded v1. Zero community issues. No adapter needed. Just download and run.

v2 is our thank you. 10x the training data, production coding capabilities, deep reasoning traces, and the same frictionless experience.

How It Was Made

Training: DeadByDawn101/unsloth-mlx — our fork with Gemma 4 support we added ourselves. FastVisionModel routes text-only training through the VLM pipeline, handling multimodal weight management automatically while LoRA touches only text transformer layers.

Fusing: Custom surgical fuse script. Dequantize each LoRA-targeted weight from the 4-bit base, add the scaled LoRA delta (lora_b.T @ lora_a.T * scale), requantize back to 4-bit. Same format, same key count, same config. No adapter needed.

Hardware: M4 Max 128GB MacBook Pro. 2 hours 4 minutes training time.

The RavenX Gemma 4 Stack

Repo	What
unsloth-mlx	Training framework — we added Gemma 4 support
mlx-gemma4	Custom model implementation + converter
ravenx-mtp-drafter	Reverse-engineered Google's hidden MTP heads for speculative decoding
ravenx-training-gym	Harbor-native security benchmark

Other Formats

Format	Link
v1 MLX (Opus only)	gemma-4-E4B-Agentic-Opus-Reasoning-GeminiCLI-mlx-4bit
v2 LoRA adapters	gemma-4-E4B-v2-Sol-FABLE5-lora
GGUF (v1)	gemma-4-E4B-Agentic-Opus-Reasoning-GeminiCLI-GGUF

About RavenX AI Labs

Security AI infrastructure company. San Jose, CA. 200K+ HF downloads. 24+ shipped models. 2 USPTO patents filed.

—USPTO #64/087,357 — Soul Infusion (identity-framed training)
—USPTO #64/104,760 — Sovereignty Chain (cryptographic model protection)

GitHub: @DeadByDawn101 | X: @RavenXllm

"We trained Gemma 4 in April. Unsloth published their guide in July. We simply do what others can't."

— RavenX AI Labs LLC, since June 2026