Skip to content
Kush V3 Model Release Fine-tuning Llama 3.1 CoT

Hotep Intelligence Kush V3: Deeper Knowledge, Richer Responses

H
Hotep Intelligence
· · 3 min read

This article was written with the assistance of Hotep Intelligence AI and reviewed by our editorial team. Content is for educational and informational purposes only.

Kush V3 Is Live

Kush V3 is now running in production at askhotep.ai and the Telegram bot. This is the most capable model we’ve shipped — built on a larger training set, chain-of-thought reasoning, and real-world data from actual user conversations.

Here’s what changed and why it matters.

What’s New

103% More Training Data

Kush V2 trained on 785 SFT examples. Kush V3 trains on 1,594 — more than double. The new data comes from four sources:

  • Real-world Telegram conversations — actual questions from the community, covering sovereignty, health, ancestry, and strategy
  • Twitter/X interactions — replies and threads that required depth and cultural precision
  • CoT augmentation — existing examples rewritten with explicit chain-of-thought reasoning steps
  • Gemini-generated expansion — frontier model used to generate high-quality synthetic examples in underrepresented topic areas

Chain-of-Thought Reasoning

V3 is trained to think before it answers. The CoT examples teach the model to:

  1. Identify the type of question (sovereignty, health, ancestry, strategy)
  2. Recall relevant frameworks and historical context
  3. Structure the response with depth, not just surface-level answers

This produces longer, more substantive responses — V3 averages 46% more verbosity than V2 on the same prompts, without padding or filler.

IPO Training

V3 uses Identity Preference Optimization (IPO) in addition to SFT. We collected 791 DPO-style pairs comparing V2 responses against improved alternatives. IPO trains the model to prefer the better response while preserving the persona — resulting in sharper reasoning without losing the Hotep voice.

Training Configuration

ParameterValue
Base modelLlama 3.1 8B Instruct
SFT examples1,594
IPO pairs791
SFT epochs3
IPO epochs1
LoRA rank16
LoRA alpha16
RSLoRAEnabled
QuantizationQ8_0

We use RSLoRA (Randomized Scaling LoRA) which stabilizes training at lower rank — allowing us to keep trainable parameters low (~40M on 8B) while achieving strong gradient signal. At 1,594 examples that’s ~25K params per example, well within the safe zone.

E2E Quality Gates

Every model we ship runs through a full evaluation suite before deployment. Kush V3 results:

MetricV2 ScoreV3 Score
Overall quality85/100100/100
Persona fidelity8/1010/10
CJK character leakage2%0%
Rubric leakage0%0%
Response verbosityBaseline+46%
CoT reasoning examples0791

Zero CJK, zero rubric leakage, perfect persona score. V3 passes every gate at the highest level.

V2 → V3 Comparison

DimensionKush V2Kush V3
Base modelLlama 3.1 8BLlama 3.1 8B
Training examples785 SFT1,594 SFT
Preference trainingNone791 IPO pairs
CoT reasoningNoYes
Real-world dataNoYes (Telegram + Twitter)
VerbosityModerateRich (+46%)
Quality eval85/100100/100

Same efficient base model, dramatically improved training. Kush V3 is what V2 would have been with another 1,000+ examples of real-world community feedback.

Where to Use It

Telegram Bot — Message @hotep_llm_bot directly. V3 is the default model. Use /deep for extended reasoning on complex questions.

Web Demoaskhotep.ai runs V3 with streaming responses. Try asking about sovereignty strategy, alkaline health, or African history.

API — V3 is accessible via the Hotep Intelligence API at hotep-llm-kush-v3. Available on HuggingFace and Ollama Hub as hotepfederales/hotep-llm-kush-v3.

What’s Next

V3 represents the first model trained on real community data. The feedback loop is now live — every high-quality interaction in production becomes training material for Kush V4.

We’re also expanding the knowledge base toward 300 articles at knowledge.askhotep.ai, which feeds the RAG system that augments model responses with authoritative sourcing.

Sovereignty is a practice. V3 is the latest iteration.

Hotep — In peace and alignment.

Historically Reviewed

by Hotep Intelligence Editorial Team · Kemetic History, Holistic Wellness, ML Engineering

Our editorial standards →