GamiWays

15.10.2026 · Research — evaluation candidate

Inworld Inference

Inworld Inference is an LLM option to evaluate in a modular cascade. Start with Gemma 4 26B A4B against the current provider; retain defaults until French evaluation is complete.

Published gains are vendor measurements, not results from our projects. Public inference is US-based; Enterprise regions and model availability require confirmation. ZDR must be activated and does not cover routed third-party providers.

The OpenAI-compatible API does not serve embeddings: RAG stays separate. First-token fallback adds its waiting time; tools and JSON outputs require validation for each model, including after fallback.

Decision Framework

Choose a voice stack from real needs: language, voice, emotional direction, RAG, latency, cost and usage type.

The simulator starts with two concrete grounds — Où est Ava ? and the Plastic Dilemma — without turning them into a recipe. Où est Ava ? prioritises emotional credibility, a French voice, controlled storyworld retrieval and avatar synchronisation; the Plastic Dilemma prioritises accessibility, pedagogical RAG, reliable feedback and a cohort-compatible cost profile. Each criterion remains reusable for another conversational experience. Results are candidates to test, not prescriptions: validate the French voice, sensitive lines, RAG corpus and total cost in actual project conditions.

🧭

Stack Decision Simulator

Profile, voice, RAG, usage and cost → ranked STT/TTS services and per-layer recommendations

1 / 11
🧩

Which project context should guide the recommendation?

Start from a concrete use, then keep every setting reusable for a different conversational experience.

On the timing of decisions

Validation phase

Prioritize iteration speed over sovereignty. Cloud APIs allow validating product hypotheses without infrastructure. Accept the technical debt of vendor dependency.

Production phase

Trade off between cost, sovereignty, and operational complexity. This is when to decide which layers deserve to be internalized, based on real data from the validation phase.

Scale phase

Architecture decisions made in the validation phase become structural constraints. Migrating a layer at scale costs 5–10× more than a well-considered initial decision.