Gamemodes is building an AI-native NPC engine that generates psychologically authentic, dynamic dialogue — running on consumer hardware with 1-bit quantized models.
For decades, NPCs have been limited to pre-written dialogue trees. Gamemodes changes that — an engine that generates contextually aware, psychologically grounded dialogue in real time, at a cost and compute footprint that works on any gaming PC.
NPCs are driven by structured psychological profiles — not random personality quirks. Every response reflects the character's internal state, emotional history, and relationship context.
Our engine uses 1-bit quantized models (~1GB VRAM) to run entirely on local hardware — no cloud dependency, no latency, no per-call API costs.
The core engine is designed for general applicability. We've proven it across three distinct game worlds — the architecture adapts to any game genre or setting.
The Gamemodes engine sits between the game and an LLM, translating game state into psychologically rich prompts and routing responses back as natural dialogue.
Model-agnostic design. Swap between any OpenAI-compatible LLM at runtime — from 1-bit models to frontier models — without changing a line of game code.
If the LLM is unavailable, the engine falls back to curated dialogue organized by psychological mode. The game never breaks.
Integrated pipeline for generating, validating, and filtering fine-tuning data — producing thousands of high-quality Alpaca-format examples.
Gamemodes operates across four interconnected projects — each proving a different dimension of the engine's capability.
Shadow City is a noir RPG designed to showcase the Gamemodes NPC engine. Set in a rain-soaked dystopian metropolis where six factions wage silent war, it demonstrates the full capability of the engine — quest systems, faction progression, mortality, and dynamic NPC-NPC interactions — all powered by psychologically grounded dialogue generation.
In development: a Skyrim Special Edition mod that replaces canned voice lines with real-time, psychologically grounded dialogue. Players type messages to NPCs via a console command; the engine generates responses through a C++ SKSE plugin communicating with a local Python dialogue server.
In development: a Fallout 4 mod using the same engine architecture — F4SE plugin captures player input, routes through a Python dialogue server, and generates wasteland-appropriate dialogue. Nine NPCs across five factions, each with full psychological profiles.
The core fine-tuning and engine project. Trains lightweight LLMs to generate psychologically authentic NPC dialogue, with a complete data pipeline from generation through deployment. Designed for general applicability across any game genre.
Gamemodes operates a complete pipeline that transforms game simulation data into production-ready dialogue models — validated, quality-filtered, and optimized for edge deployment.
Game simulations produce thousands of NPC interactions. A larger LLM generates training examples across categories — dialogue, faction voice, affinity scenarios, moral dilemmas, and more.
Every example is validated for schema compliance, field completeness, and distribution balance. Duplicates are removed automatically.
An independent LLM scores each example across five dimensions: psychological authenticity, state visibility, voice consistency, genre accuracy, and relationship modeling. Only high-quality data advances.
Clean data is converted to Alpaca format (instruction/input/output) with multiple instruction templates and controlled augmentation to increase effective dataset diversity.
QLoRA fine-tuning on meta-llama-3.2-1b-instruct — proven settings with rank=16, adamw_8bit. Exportable to GGUF q4_k_m for local inference via llama.cpp.
Models are served via llama.cpp's OpenAI-compatible API. The engine routes requests between models based on complexity — Bonsai 8B for nuanced scenes, Llama 1B for simple exchanges.
Our roadmap centers on 1-bit quantized models — leveraging Microsoft's BitNet research and the Bonsai model family to deliver frontier-quality NPC dialogue at a fraction of the compute cost.
| Model | Parameters | Quantization | VRAM | Role | Status |
|---|---|---|---|---|---|
| Llama 3.2 1B | 1.2B | Q4_K_M | ~0.5 GB | Simple exchanges, fallback | Active |
| Bonsai 8B | 8.2B | Q1_0 (1-bit) | ~1.0 GB | Complex dialogue, nuanced scenes | Active |
| Bonsai 4B | 4B | 1-bit | ~0.7 GB | Mid-tier routing target | Planned |
| BitNet (Microsoft) | Variable | 1.58-bit | TBD | Next-gen baseline for future development | Research |
Each game produces its own fine-tuned adapter — trained on that game's specific dialogue patterns, lore vocabulary, and character archetypes. One base model, many game-specific variants.
The engine architecture is game-agnostic. Proven across noir RPG (Shadow City), fantasy (Skyrim), and post-apocalyptic (Fallout 4) — any genre can plug in.
1-bit models deliver 8B+ parameter quality at ~1GB VRAM. This means NPC dialogue generation runs on the player's existing GPU — no cloud costs, no latency, full privacy.
Gamemodes welcomes contributors across every discipline — from game modders to ML researchers.
Help expand NPC profiles, write training scenarios, or build integrations for new games. Familiarity with Skyrim/Fallout 4 modding is a plus.
Work on the dialogue server, training pipeline, quality filtering, or simulation engine. Flask, data processing, and API design experience welcome.
Contribute to the SKSE/F4SE plugin layer — NPC resolution, HTTP clients, subtitle display, or Papyrus integration. CommonLibSSE-NG experience helpful.
Help optimize fine-tuning strategies, explore 1-bit quantization approaches, improve quality scoring, or design new training data categories.
Craft training scenarios, define psychological profiles, design archetypes, or help shape the narrative systems. Understanding of character psychology valued.
Run simulations, validate training data quality, test dialogue outputs across game scenarios, or help build automated evaluation frameworks.
Whether you want to contribute, invest, or just learn more — we'd love to hear from you.
✉ gamemodes.dev@gmail.com