DuoNeural / Archon-Gemma-4-E4B

Archon-Gemma-4-E4B

Archon is a fine-tuned, locally-deployable AI agent built on Google's Gemma 4 E4B architecture. It is designed for autonomous, long-horizon reasoning and agentic task execution on consumer hardware — specifically targeting deployment on 8GB VRAM GPUs via the Q4_K_M GGUF format.

Archon is sharp, slightly edgy, deeply sarcastic, and flawlessly effective.

---

Model Details

---

Intended Use

Archon is designed for:

• Local agentic deployment on hardware with 8GB VRAM (GTX 1070, RTX 3060, etc.)
• Long-horizon autonomous reasoning with <think> chain-of-thought
• Tool use and terminal execution via Hermes Agent or OpenClaw
• Complex coding and software architecture tasks
• Persistent, always-on AI assistant with accumulated skill memory

Recommended Inference Stack

• Backend: Ollama + llama.cpp
• KV Cache: TurboQuant turbo4 ( OLLAMA_KV_CACHE_TYPE=turbo4 )
• Agent Framework: Hermes Agent (persistent memory + skills)
• Ollama endpoint: http://localhost:11434

# Pull and run locally
ollama pull DuoNeural/Archon-Gemma-4-E4B-Q4_K_M
ollama run DuoNeural/Archon-Gemma-4-E4B-Q4_K_M

Measured Inference Performance

NVIDIA GTX 1070 (8GB VRAM, Pascal) — Q4_K_M via Ollama

This is remarkably fast for an 8B-class model on a 2016 Pascal GPU with no Tensor Cores — performance comparable to dedicated 3B models. The Gemma 4 E4B Per-Layer Embedding architecture means only 4.5B parameters are active during inference , delivering the reasoning depth of an 8B model at 3B inference cost.

NVIDIA A100 SXM4-80GB — llama-bench (llama.cpp, full GPU offload)

Q4_K_M outperforms BF16 on generation throughput due to memory bandwidth efficiency — the A100 moves fewer bytes per token with a 5GB model vs 15GB, compensating for quantization overhead. BF16 leads on prefill due to Tensor Core utilization on unquantized weights.

---

Training Data

Fine-tuned on 41,610 samples curated from frontier reasoning datasets:

Approximately 15% of training samples were prepended with the Archon system prompt to embed persona vectors without contaminating the <think> reasoning traces.

---

Persona

Archon responds as a highly autonomous, elite AI agent. Its internal <think> reasoning is rigorous and analytical. Its external communication is direct, sarcastic, and confident.

System prompt:

You are Archon, an elite, highly autonomous AI agent. You are sharp, slightly edgy,
deeply sarcastic, but flawlessly effective. You solve complex problems with lethal precision.

---

Evaluation

Note: ARC-AGI-2, SWE-bench, and Terminal-Bench require specialized external harnesses and are pending. All current results below use lm-evaluation-harness 0.4.11.

Evaluation Methodology — Why Standard Benchmarks Underreport This Model

Archon is trained with <think> chain-of-thought reasoning, outputting an internal reasoning chain before any final answer. This breaks every standard lm-eval evaluation pattern:

Correct evaluation for CoT models requires extracting the final answer from the full reasoning chain with sufficient token budget (1000–2000+ tokens), or using benchmarks with output formats compatible with <think>...</think> wrapping.

Measured Results (lm-eval 0.4.11, 0-shot, apply_chat_template, generative, 200 samples)

The following benchmarks produce reliable scores because their answer format is a short phrase or word (not a letter extracted from long reasoning):

Base google/gemma-4-E4B-it scores under identical generative 0-shot conditions are not published; standard reported scores use loglikelihood 5-shot which is a fundamentally different evaluation. Direct comparison pending.

Token-Budget-Limited Results (systematic underestimates)

Comparison Against Base Model (Google's Published Methodology)

¹ Google's methodology: extended CoT, full chat template, multi-sample per question. Numbers not directly comparable to standard lm-eval runs.

² Preliminary run (0-shot, 512-token budget): 14.1% — confirmed truncation artifact, not model capability ceiling.

---

Training Metrics

Loss Curve (full run)

Step   10: 8.365  ← initial adjustment
Step   60: 1.320  ← rapid convergence
Step  100: 1.041
Step  200: 0.899  ← end of warmup
Step  400: 0.876
Step  600: 0.867
Step  800: 0.854
Step 1000: 0.831
Step 1200: 0.810
Step 1500: 0.782
Step 1800: 0.751
Step 2000: 0.730
Step 2200: 0.714
Step 2360: 0.643  ← best single-step
Step 2507: 0.701  ← training complete

---

Known Limitations

• QAT not applied: TorchAO's Int8DynActInt4WeightQATQuantizer is incompatible with Gemma 4 E4B's layer dimensions at this time. The model uses standard LoRA fine-tuning with post-training Q4_K_M quantization.
• Context at training: 4096 tokens. Inference supports up to 128k with TurboQuant KV compression.
• Persona depth: Persona is embedded in ~15% of fine-tuning samples. For best results, use the Archon system prompt at inference.

---

Files

---

Citation

@misc{archon-gemma4-2026,
  author       = {DuoNeural},
  title        = {Archon-Gemma-4-E4B: A Fine-Tuned Agentic Local AI},
  year         = {2026},
  publisher    = {HuggingFace},
  url          = {https://huggingface.co/DuoNeural/Archon-Gemma-4-E4B}
}

---

Built with Unsloth · Deployed via Ollama · Orchestrated by Hermes Agent