NIA is a 350-billion-parameter Mixture-of-Experts model we built from zero on a custom narrative corpus anchored by 155,000+ licensed novels and other curated sources. The full STT → LLM → TTS path runs end-to-end on bare-metal Blackwell B200s, walled behind our own FastAPI. One engineered system, not orchestrated agents. No OpenAI, no Anthropic, no Gemini anywhere in the hot path. We optimized for emotional intelligence - the axis the frontier labs treat as an afterthought.
Three layers, one path, one piece of hardware. Everything between the player's microphone and the AI-generated audio output is code and weights that we own and run.
NIA isn't "agentic AI" - not dozens of LLM calls stitched together with prompts, fallback chains, and hope. It's a single purpose-built system: model, custom tokenizer, state machine, serving stack - all designed and built as one piece. The output isn't a stitched approximation of coherence. It's coherent by construction.
The deeper technical story - SGLang fork, EAGLE speculation, FlashInfer version hell, the Blackwell-specific kernel discoveries - is documented in the founder's Skunkworks deep-dive.
Four numbers worth memorizing. The rest live in the technical write-up.
Read more about our technical struggles - the Skunkworks article walks through the full inference-stack journey, including the +23.7% TTFT win from MoE shared-expert fusion and why FP8 KV cache turned out to be slower than BF16 on Blackwell.
Frontier LLMs - Claude Opus, GPT-class, Gemini Ultra - compete on MMLU, HumanEval, GPQA. Pure IQ benchmarks: reasoning, code, math, world knowledge.
NIA was built against a different objective entirely - from corpus and tokenizer to architecture and serving. Narrative coherence. Character continuity. Emotional pacing. Tonal control across long arcs. These don't exist as public benchmarks because the field treats emotional intelligence as an emergent downstream property of scale. We treated it as a primary objective end-to-end.
The proof is in deployment, not in argument: our first B2B customer (in HR) compared NIA head-to-head against the frontier models and chose NIA - explicitly on the EQ axis, where SOTA's deep reasoning didn't translate to the human-facing scenarios they needed to handle.
NIA wins when the use case is: an extended human-facing conversation, with consistent persona and emotional state, where the agent has to respond like a person who has been listening, not retrieve and reason. Storytelling. HR. Coaching. Training. Anywhere the conversation matters more than the answer.
What every other AI system calls "memory" today is RAG - retrieve relevant chunks of past context, paste them back into the prompt, hope the model uses them. The model itself stays stateless between turns. Memory is bolted on, not architectural.
Liquid Memory is different. NIA tracks emotional history, broken promises, character relationships, and structural narrative state across tens of millions of tokens within a single session - architecturally, not by retrieval. Betray an NPC in chapter one and they remember you in chapter five. Not because the right chunk got pulled back into the prompt. Because the model carries the state.
Liquid Memory is backed by a pseudo-state non-deterministic machine paired with a custom tokenizer purpose-built for narrative. Standard frontier-LLM tokenizers are optimized to encode literal text efficiently. Ours encodes what's said between the lines - subtext, implication, broken promises, emotional weight. The state machine reads those signals and updates persistent narrative state turn by turn.
This is also the single biggest reason a generic LLM cannot do what NIA does - even at frontier scale. SOTA models were trained against tokenizers optimized for different objectives; there's no surface for our state machine to attach to. RAG-class systems patch over the absence of long-form coherence; they don't produce it.
Max walked through Liquid Memory and the broader design choices in a long-form exclusive with Korea Game Desk - their writeup is the closest thing to a public architecture note we've put out.
Every AI gaming, narrative, or character-AI company you'll see this year is built on third-party APIs - OpenAI, Anthropic, Google. Their margin gets taxed on every single user interaction. Their cost-per-interaction floor is set by someone else's API price list.
NIA on bare metal is the opposite. We pay for compute. The per-token margin stays inside the business.
Even if Claude Opus could meet our latency budget on this workload - it can't - the per-token cost would be roughly 80× higher than what we pay on bare metal. The use case literally doesn't exist on a frontier-API stack. That's not an opinion about our moat. That's the cost sheet.
Demo period averaged 3.3% server capacity used across 5× B200. We can roughly 30× the concurrent player count on the same hardware before provisioning new GPUs. Cost per concurrent user at this scale is below anything running on rented inference.
Primary focus is the Omea consumer launch on Steam and App Store - that's where the company's energy is right now. B2B is real and earning revenue, but it grows alongside the consumer launch rather than ahead of it.
Hosted on our bare metal, walled behind FastAPI. Per-token pricing, contract per partner. Not yet open / self-serve - by design: capacity is managed, partners are vetted, the model is not a commodity endpoint.
If you're a business angel evaluating an investment, the full pitch and round terms live in the investor brief. If you want a deeper technical conversation, hit reply.