Introducing Inworld Inference
Inworld launches Inference, an inference API for consumer AI apps that optimizes open models such as Gemma 4 26B. It claims first tokens up to 7.8x faster at p95 than the median OpenRouter provider and input prices up to 48% lower.
What the article says
On October 14, 2026, Inworld announced Inworld Inference, an LLM inference API for consumer AI apps such as companions, tutors and coaching, open to all developers. It targets specific needs: fast replies in conversation, long prompts with short outputs, many calls per turn, low per-user cost and high reliability.
It offers optimized builds of open models (Gemma 4, GLM, DeepSeek, MiMo), realtime decoding where a draft model proposes tokens that the full model verifies, KV-aware routing, four service tiers (Flex, Default, Priority, Fast), automatic failover to fallback models, an OpenAI-compatible API and dedicated deployments. By the company's own measurements, Gemma 4 26B shows a p95 time to first token of 0.31 s versus 2.42 s for the median OpenRouter provider, at $0.07 per million input tokens.
Read the official article AnnuIA profile
Summary written by AnnuIA in its own words from the official page of Inworld AI. The content and rights belong to the publisher: please read the original article.