NVIDIA's nemotron-cascade-2 is now the primary LLM across the entire Jefe ecosystem — JefeCode, all three Discord/Twitch bots, and all public-facing website chatbots. We deployed and benchmarked three NVIDIA inference backends (vLLM, TensorRT-LLM, NIM), shipped the Adventure GUI with a working AI dungeon master, hardened FreeChat with six security fixes, and overhauled the Inception deck with real benchmark data. Nineteen agents, one session, zero aspirational claims remaining.

Nemotron Takes Over

Every LLM-powered service in the ecosystem now runs on NVIDIA's nemotron-cascade-2 — a 30B MoE model that only activates 3B parameters per token. JefeCode serves it via vLLM (NVFP4 quantization, 238 tok/s), while bots and chatbots route through Ollama (259 tok/s). The bots were all on OpenAI's gpt-4o-mini — switching them off was both a cost savings and an NVIDIA alignment move. Website chatbots also gained multi-turn conversation context (last 10 messages) so follow-up questions actually work now.

Three Inference Backends, Real Numbers

We deployed TensorRT-LLM 1.3.0 and NVIDIA NIM on the 96GB VRAM Blackwell GPU and benchmarked them against our existing vLLM setup. TRT-LLM matched vLLM at 238 tok/s for single-request throughput (with 15-30% advantage under batch load). NIM ran Llama 3.1 8B at 139.3 tok/s after we manually forced the SM120 profile (auto-detection doesn't recognize Blackwell yet). vLLM stays primary, but having three NVIDIA backends evaluated with real data is exactly the story the Inception deck needs.

Adventure GUI: From Mock to Real DM

The Adventure GUI frontend was deployed with mock responses hardcoded to true, which meant the DM was cycling through four canned strings regardless of player input. We fixed the mock flag, added think-tag filtering for Qwen3/3.5 models, and reinforced player agency in the system prompt. The DM now actually responds to what you say. Also fixed TTS on Titan (three bugs: tkinter import crash in headless Docker, missing NVRTC for GPU inference, and a missing spaCy phonemizer model). The full 6-stage GygaxBot archival pipeline is operational.

FreeChat Security Hardening

The security engineer found six vulnerabilities: PBKDF2 iterations at 100K instead of 600K (across four codepaths), an SSRF via DNS rebinding in the link preview endpoint, OAuth CSRF bypass when cookies are absent, no rate limiting on token refresh, stored XSS via filename injection, and WebSocket reactions that didn't verify room membership. All six fixed, all tests passing. Existing users will need a password change to benefit from the new iteration count.

Container Security and Infrastructure

Created the jefedata shared group (GID 1500) on Titan, locked all .env files to 600 permissions, applied setgid to shared directories, and reverted every chmod 777 workaround. Zero world-writable directories remain. Adjusted Prometheus alerts from an aggressive 4GB per-container threshold to 80% system memory with a 20GB per-container cap — inference containers hitting 8-24GB RAM is normal, not an emergency. Scrubbed 16 devlog posts of internal infrastructure details.

Inception Deck: Zero Aspirational Claims

The NVIDIA Inception deck went through multiple revision rounds. Every claim is now backed by deployed systems and real benchmarks. Nemotron-cascade-2 is the headline (238 tok/s on Blackwell), a four-backend benchmark comparison table shows systematic evaluation, Adventure GUI has the "own your experience" pitch (zero subscription, unlimited usage, your GPU your game), GygaxBot's archival pipeline demonstrates D&D ecosystem depth, and TRT-LLM/NIM language reflects actual deployment — not planned evaluation.

New Agents and Design Specs

Created two new agents: jefeai-lab for model research, evaluation, and deployment (separating it from jefeai-engineer's overloaded plate), and adventure-designer for Adventure GUI product design, narrative quality, and game mechanics. The adventure-designer produced two comprehensive design specs: a character creation and sheet system (4 stat generation methods, DM/Player visibility mode, skill picker) and a Manual DM mode where a human runs the campaign with AI tools as assistants rather than replacements.

What's Next

• NVIDIA Inception submission — final review, demo video, screenshots
• Adventure GUI character system (Phase 1 implementation from design spec)
• Manual DM mode implementation
• Nemotron-3-Super NVFP4 evaluation via vLLM
• NeMo Guardrails for JefeHealth, NeMo-RL for LoRA fine-tuning
• ComfyUI --vram-timeout open-source contribution