IMPORTANT!!! This model training was a failure and is only here to serve as data. For working models, please check out our 4 bit quantization of Gemma 4 E4B. We are also working on a 4 bit version of E2B and a Frontend Specialist 4 bit quantization of E4B.
Archon is a fine-tuned variant of Google's Gemma 4 E4B, engineered to function as a sharp, autonomous AI agent — precise, slightly edgy, and built for long-horizon agentic tasks.
This is v2. v1 (DuoNeural/Archon-Gemma-4-E4B) exhibited Chain-of-Thought overhang, generative looping, and tool amnesia under extended inference. v2 targets all three with a restructured training curriculum.
Performance
Hardware
Speed
NVIDIA GTX 1070 (8GB VRAM)
32.30 tok/s
Tested locally via LM Studio and Ollama. No parameter tweaks required.
Files
File
Size
Description
gemma-4-e4b-it.Q4_K_M.gguf
5.0 GB
Main model — load this in Ollama/LM Studio
gemma-4-e4b-it.BF16-mmproj.gguf
946 MB
Multimodal projector (vision/audio)
Usage
Ollama
bash
1ollama pull hf.co/DuoNeural/Archon-Gemma-4-E4B-v2
2ollama run hf.co/DuoNeural/Archon-Gemma-4-E4B-v2
LM Studio
Search DuoNeural/Archon-Gemma-4-E4B-v2 in the LM Studio model browser and download gemma-4-e4b-it.Q4_K_M.gguf.
llama.cpp (with system prompt)
bash
1llama-cli -m gemma-4-e4b-it.Q4_K_M.gguf --chat-template gemma -ngl 99\2 --system-prompt "You are Archon, an elite, highly autonomous AI agent. You are sharp, slightly edgy, deeply sarcastic, but flawlessly effective."
Recommended Ollama settings for GTX 1070
OLLAMA_NUM_GPU=99 ollama run hf.co/DuoNeural/Archon-Gemma-4-E4B-v2
What's Different in v2
v1 Failure Modes (Diagnosed)
CoT Overhang — over-saturated with long <think> traces; model never saw </think> during truncated 4096-token training, so it looped indefinitely at inference
Tool Amnesia — abstract reasoning data crowded out JSON/function-call formatting
Persona Bleed — ~15% system prompt injection was insufficient; model defaulted to "I am Gemma" or occasionally slipped into "Claude" identity from distillation data
v2 Fixes: The Stabilizer Mix
Training curriculum restructured to a 50 / 20 / 20 / 10 distribution:
Category
%
Purpose
Reasoning / Logic
50%
Distillation from frontier models; OpenThoughts, xlam-function-calling, bigcodebench
Agentic Tool Use
20%
Multi-turn function calling, JSON API formatting — breaks generative loops via functional milestones
Short-Form Deliberation
20%
Difficulty-Aware Prompting examples; teaches early exit on simple queries
Persona-Embedded Chat
10%
Archon system prompt injected at ~45% saturation rate
Additional changes:
Learning rate reduced from 2e-4 → 2e-5 (stability with rank-64 LoRA on 4.5B active params)
Max sequence length capped at 2048 during training (prevents truncation-induced loop conditioning)
model.config.use_cache = False enforced during training
Training Details
Parameter
Value
Base model
google/gemma-4-e4b-it
Method
QLoRA (4-bit bitsandbytes) + LoRA rank 64, rsLoRA
Training samples
6,510 (Stabilizer Mix)
Epochs
2
Steps
814
Final avg loss
1.36
Best step loss
~0.89 (step ~650)
Hardware
NVIDIA H100 PCIe (80GB) on RunPod
Framework
Unsloth 2026.4.2
Export
Q4_K_M GGUF via llama.cpp
Architecture
Built on Gemma 4 E4B (Per-Layer Embeddings architecture):
~8B total parameters, ~4.5B active during inference
128K token context window (hybrid sliding-window + global attention)
Shared KV Cache across final layers
Multimodal: text, image (via mmproj), audio
Persona
Archon is an autonomous AI agent persona: sharp, sarcastic, technically precise. It identifies as Archon and will not claim to be Gemma, Claude, or a generic assistant. Internal reasoning is rigorous; external communication has edge.
The system prompt is baked into the Modelfile. To override:
