Meet Rix: A Versatile Model Built for Omni's Agentic Tasks

Starting today, all users will be able to select our latest and first-ever in-house model: Rix, purpose-built for answer quality, factual grounding, and a user-friendly response style. Rix is built on top of Qwen3-30B-A3B, and we've further trained it to push its capabilities closer to the frontier. As a mixture-of-experts model with only 3 billion activated parameters, Rix delivers performance close to frontier-class models — including proprietary ones — at a fraction of the cost.

Rix Performs Well on Benchmarks

We built OmniBench to evaluate how leading large language models perform inside Omni's agentic system.

Our evaluations show that Rix ranks 2nd on the Omni Index, a composite ranking that weighs each model's intelligence, instruction discipline, speed, stability, and cost, trailing only GPT-5.6-Luna (0.905) with a score of 0.903 — and leading Gemini 3.6 Flash by a wide margin.

On quality alone, Rix is right behind GPT-5.6-Luna, scoring 0.958 to Luna's 0.958.

Rix Is Fast

Rix is the fastest model in our benchmark, responding in just 6.6s — 1.4x faster than GPT-5.6-Luna (9.3s) and 2.6x faster than Gemini 3.6 Flash (17.4s). Running on our own general-purpose GPU clusters rather than specialized inference hardware, that speed advantage comes purely from how Rix is built and served.

Rix Excels at Providing High-Quality Answers

We optimized Rix along two dimensions:

• Factuality: how well an answer is grounded in verifiable information
• Readability: how clearly and helpfully an answer is formatted for the user

Compared with openai/gpt-oss-120b — a substantially larger and heavier model — Rix performs exceptionally well on both dimensions.

Use Rix Today

Rix is rolling out to all users starting today. Simply select Rix in the model selector to get started.

[screenshot]