Inworld Launches Inworld Inference, an API It Says Is Built for Consumer AI Apps

Inworld, a company that builds AI for consumer apps, has opened Inworld Inference to all developers. The service runs open AI models for apps such as AI tutors, coaches and companions.

Inworld announced the launch in a blog post on 14 October 2026. It calls the service “the first inference service designed for consumer AI”, a claim that has not been independently checked. Inference is the work of running a trained AI model to produce answers.

Why Inworld says consumer AI is different

Inworld argues that most AI providers focus on business customers and coding, while consumer apps need fast replies to long conversations at low cost.

“Today only 2% of US households, and 3% of US consumers, spend anything on AI,” Inworld writes, citing figures from the venture capital firm a16z using PNC data.

What it offers

According to Inworld, the service:

• hosts open models including Gemma 4, GLM, DeepSeek and MiMo, each with its own tuned serving setup;
• has no rate limits and lets developers choose a service tier per request;
• can route across models with automatic failover, or run a customer’s own model on dedicated hardware; and
• uses an OpenAI-compatible API, so developers can switch with few code changes.

What Inworld says about speed and price

All of the following are Inworld’s own measurements and comparisons:

• Speed: on Google’s Gemma 4 26B model, Inworld says its first reply token arrives up to 7.8 times faster than from the median provider on OpenRouter, an AI model marketplace (0.31 seconds against 2.42 seconds at the slowest 5% of requests).
• Price: Inworld says it charges $0.07 per million input tokens for Gemma 4 26B, nearly 40% below the median provider that doesn’t train on customers’ data.
• Efficiency: it says the same hardware now handles about 1.8 times as many requests as in September, and up to 90% of a prompt can come from its cache, at up to 99% off.
• Reliability: it says 99.999% of more than 20 million requests to the GLM 5.3 Flash model in 30 days completed without an error.
• Scale: Inworld says it serves “tens of trillions of tokens a month, 25 times what we served in March”.

Inworld itself notes that its speed comparisons are “Snapshots of each source, not a matched workload benchmark”. Its own numbers were measured at its router, while rivals’ figures come from OpenRouter’s measurements.

What this means

• Consumer AI apps are a new battleground. Inworld is betting on apps people use in daily life, not office tools.
• The numbers are Inworld’s. No independent benchmark has confirmed its speed or price claims.
• Developers can try it easily. The OpenAI-compatible API lowers the cost of switching.

Sources

1. Inworld, “Introducing Inworld Inference”, 14 October 2026. https://inworld.ai/blog/introducing-inworld-inference

Related stories

• Watching AI Agents From the Inside, and Rewiring the AI Data Centre
• Hugging Face Launches Llama Windows, a Free App That Runs Open AI Models From Your Windows 11 Tray
• Fish Audio Launches Drama 3, a Text-to-Speech Model You Direct in Plain Language
• Fireworks AI Confirms Misuse of Internal Credentials and Capture of Customers’ Stored API Keys and Tokens; It Reports No Evidence So Far of Wider Data Access