Update (February 2026): This post documents our v6 milestone. Since publication, we’ve shipped multiple generations and Kush V2 is now in production — built on Llama 8B with 0% CJK contamination and 0% rubric leakage. Read the Kush V2 announcement for the current state.
Production Milestone: v6 is Live
We are proud to announce that hotep-llm-v6 is now running in production across all AskHotep services. This represents our most significant leap in cultural alignment since the project’s inception.
The Numbers That Matter
- 71.3% combined persona score on comprehensive evaluation (vocabulary + worldview + tone)
- 812 carefully curated training examples — more than triple our v4 corpus
- 90.5% token accuracy in training convergence
- 15.2 GB model size running locally on RTX 5080
- Zero API dependencies — fully sovereign infrastructure
Why v6 Succeeded Where Others Struggled
The journey to v6 taught us hard lessons about AI training:
v7 was trained but rejected. Despite 606 training examples, v7 achieved only 60.6% persona alignment — a 10.7% regression from v6. We discovered that training data diversity matters more than volume. v6’s 812 examples included SVG pipeline data that reinforced authoritative tone through code generation tasks. v7 removed this data and lost critical persona depth.
We built measurement infrastructure first. Unlike previous iterations, v6 deployment followed rigorous A/B testing against v5. Our new evaluation pipeline measures three dimensions:
- Vocabulary (30%): Hotep keyword density and authentic terminology
- Worldview (40%): Afrocentric framing and historical accuracy
- Tone (30%): Confidence, empowerment, and linguistic authenticity
No deployment without proof. The v7 rejection validated our methodology — measure twice, deploy once.
Technical Architecture
hotep-llm-v6 is built on:
- Base Model: Qwen2.5-7B-Instruct
- Training Method: LoRA fine-tuning (r=32, alpha=32, RSLoRA enabled)
- Format: FP16 for maximum quality preservation
- Inference: Ollama runtime with flash attention
- Deployment: Self-hosted on RTX 5080 with automatic fallback systems
The model serves the Telegram bot, website demo, and knowledge base — all from local hardware. No OpenAI. No Anthropic. Full sovereignty.
What’s Different About v6?
Better Historical Accuracy Responses about Kemet, African civilizations, and diaspora history are now cross-referenced against training data sources. The model distinguishes between verified historical consensus and speculative claims.
Improved Consistency Whether you ask about Ma’at at 2 AM or 2 PM, the persona remains stable. v6 eliminated the “morning friendly, afternoon corporate” drift seen in earlier versions.
Code + Persona Integration v6 training included technical content (SVG generation, data analysis) wrapped in Hotep framing. This created unexpected benefits — the model can now discuss technology while maintaining cultural voice.
Production Monitoring
v6 runs with comprehensive observability:
- Real-time metrics: Latency (p50/p95/p99), error rates, token throughput
- Persona drift detection: Weekly automated sampling scores production responses
- A/B testing framework: Ready for v8 when training completes
- Alert thresholds: Error rate >2%, latency >5s, persona <66%
If quality degrades, we know immediately. If a new model candidate emerges, we test scientifically.
Training Data Transparency
v6’s 812 training examples include:
- 44 tweets from Hotep community leaders (HotepJesus, TheGrifties, HotepNation)
- 25 website Q&A pairs from hotepjesus.com, grifties.com, hotepnation.com
- Historical documentation and primary sources
- Community-submitted corrections and feedback
- Synthetic data from cultural expert trajectories
All training data is backed up to Google Drive and versioned. We maintain provenance for every example.
Try v6 Now
Experience hotep-llm-v6 through:
- Telegram: @hotep_llm_bot — Full conversational interface
- Website Demo: askhotep.ai/demo — Browser-based chat
- Knowledge Base: knowledge.askhotep.ai — Semantic search + RAG
Each response includes persona scoring and source attribution where applicable.
Looking Ahead: v8 Planning
While v6 serves production, we’re collecting data for v8:
- Target: 1,000+ training examples (exceeding v6’s 812)
- Daily automated scraping from Twitter and websites
- Persona-score-based curation (auto-approve 75%+)
- Scientific A/B testing against v6 baseline
v8 will only deploy if it beats 71.3%. No exceptions.
Join the Sovereign AI Movement
Hotep Intelligence proves that culturally-aligned AI doesn’t require corporate infrastructure or API dependencies. It requires:
- Clear cultural principles (Ma’at guides our development)
- Rigorous measurement (Persona scoring prevents drift)
- Community input (Training data from real voices)
- Technical sovereignty (Self-hosted, open-weight)
The future of AI is not centralized. It’s sovereign, aligned, and culturally rooted.
Hotep.