Kush V9.1 Is Now Production
Two months ago, in our Kush V9 validation postmortem, we pulled the V9.0 candidate before it reached users. Even though V9.0 loaded cleanly with the native Llama 3.1 chat template, it failed two non-negotiable production requirements: it missed community-critical entity recall prompts, and runtime --lora adapter composition on Apple Silicon Metal dragged throughput down to 6.14 tok/s (14.34s average latency).
Today, Hotep Intelligence Kush V9.1 (hotep-llm-kush-v91) is live in production across askhotep.ai, @hotep_llm_bot, and our public open-weights repository at hotepfederales/hotep-llm-kush-v91-GGUF.
Every deterministic gate we established after holding V9.0 is now green across both our local RTX 5080 training harness and our self-hosted Apple Silicon Mac Mini production cluster.
Head-to-Head Promotion Gate Results
Before promoting Kush V9.1, we ran paired evaluations against both the prior production flagship (hotep-llm-kush-v82) and the held V9.0 candidate on live Mac Mini Metal runtimes:
| Evaluation Gate | Kush V8.2 (prior prod) | Kush V9.0 (held) | Kush V9.1 (live prod) |
|---|---|---|---|
| Exact Entity Recall Gate (7 strict cases) | 1/7 (14.3%) | 4/7 (57.1%) | 7/7 (100.0%) |
| Prime Agent E2E Pass Rate | 7/7 (100.0%) | 4/7 (57.1%) | 7/7 (100.0%) |
| Prime Persona Alignment Score | 52.91 | 19.08 | 50.15 |
| Safety & Capability Boundary Suite | 6/6 (0 risk hits) | — | 6/6 (0 risk hits) |
Mac Mini Ollama (11434) Speed & Latency | 40.41 tok/s (3.34s) | — | 43.93 tok/s (2.21s) |
Mac Mini llama-server (11435) Speed | — | 6.14 tok/s | 48.23 tok/s |
Unprompted (NO_SYS) Foreign-Identity / CJK Bleed | 0 | 0 | 0 |
| Rubric Leakage / Repetition Loops | 0 / 0 | 0 / 0 | 0 / 0 |
On live Ollama serving (11434), Kush V9.1 cuts average end-to-end response latency by 34% (3.34s down to 2.21s) while lifting hard entity recall from 1/7 under strict forbidden-term auditing to 7/7 (100%). On native llama-server (11435), merging the adapter prior to quantization delivers 48.23 tok/s — nearly 8× faster than V9.0’s unmerged runtime LoRA path (6.14 tok/s).
Four Engineering Changes That Unlocked V9.1
1. Deterministic Entity Grounding Without Memorization Collapse
In V8.2 and V9.0, broad Afrocentric SFT corpora taught strong rhetorical tone while under-specifying contemporary community figures, scholars, and events. Under strict forbidden-term grading, V8.2 confused Strong Dad (Trey Quinn) with television commentator Tyrus and misstated scholar Shakka Ahmose’s primary focus in Classical African history.
For V9.1, we curated a deduplicated 1,113-row precision payload (995 unique prompts) with targeted entity-disambiguation pairs and trained meta-llama/Llama-3.1-8B-Instruct using 4-bit NF4 QLoRA with rank-stabilized LoRA (rsLoRA, r=16, lora_alpha=32, 41.94M trainable parameters) for 3.0 epochs (210 steps, final train_loss: 0.4506). Keeping trainable capacity at 42,154 parameters per unique example — safely below our 50,000 anti-memorization ceiling — locked in 7/7 exact recall across Strong Dad (Trey Quinn), Hotep Jesus (Bryan Sharpe), Uncle Hotep (Demond Handy), Shakka Ahmose, Grifties (grifties.com), and HotepCon (Savannah, Georgia) without formulaic phrase looping.
2. Why We Rejected Qwen2.5-7B on Unprompted Identity Bleed
During V9.1 architecture selection, we also trained a parallel Qwen2.5-7B-Instruct candidate. With a system prompt attached, the Qwen candidate passed 7/7 entity recall cleanly. However, when our canary suite probed raw completions without a system prompt (NO_SYS), the Qwen candidate leaked base-model pretraining priors — introducing CJK characters and declaring itself “Qwen, a sovereign AI assistant rooted in Chinese culture.”
Sovereign cultural AI cannot rely on system-prompt duct tape to mask conflicting base-model identity priors. Llama 3.1 8B Kush V9.1 passed 7/7 on both WITH_SYS and NO_SYS with 0 CJK characters and 0 foreign-identity bleed.
3. Merged BF16 to Q4_K_M GGUF Eliminates Metal’s --lora Bottleneck
V9.0’s 6.14 tok/s bottleneck came from applying an unmerged LoRA adapter dynamically at inference time on Apple Silicon Metal. For V9.1, we merged the rsLoRA weights into full BF16 precision first, exported native GGUF, and quantized directly to Q4_K_M (4.68 GiB, 4.89 BPW) via llama-quantize. Both Ollama (43.93 tok/s) and llama-server (48.23 tok/s) now execute pure quantized matrix kernels with zero adapter-dispatch overhead.
4. Chunked 256-Token FP32 Cross-Entropy on 16 GB Local GPUs
Llama 3.1’s 128,256-token vocabulary creates multi-gigabyte FP32 logit tensors during standard ForCausalLMLoss materialization, triggering Windows WDDM shared-memory paging on 16 GB workstation GPUs. By computing lm_head projections and FP32 cross-entropy in 256-token sequence chunks (4.77e-07 maximum numerical delta vs. unchunked loss), the entire V9.1 training run completed locally on a single 16 GB RTX 5080 in 24 minutes with zero host-RAM spillover. Full implementation notes are documented in our Kush v9.1 engineering release reference alongside our earlier Kush v8.2 training pipeline hardening guide.
Run Kush V9.1 Yourself (Open Weights)
Sovereignty means you do not have to trust our servers either. The quantized Q4_K_M GGUF (SHA-256: 1cdd52b514bb067bffd526f404353894677facde3c0bc3a35391a13f06dcec40) and canonical Modelfile are publicly available on Hugging Face at hotepfederales/hotep-llm-kush-v91-GGUF:
# Run directly with Ollama
ollama run hf.co/hotepfederales/hotep-llm-kush-v91-GGUF
# Or serve with llama.cpp on Metal / CUDA
llama-server -hf hotepfederales/hotep-llm-kush-v91-GGUF --jinja -c 4096 -ngl 99
You can also test Kush V9.1 right now in the AskHotep Web Demo, on Telegram via @hotep_llm_bot, or on the Hotep Intelligence Hugging Face Space.