技術·the technology
rikAI runs a translator and a reasoning model on the phone in your hand. This page is how they were trained, what they score, and how they work together.
On the device
Translator
Turns the sentence in front of you into English — and when the app has already pinned down what a word means in this sentence, the translator is required to honor it. That per-word glossary control is trained in, not bolted on.
Reasoning
Answers your questions about the sentence — briefed first with the dictionary entries and grammar points the app matched, so it explains from facts instead of vibes. Fine-tuned for exactly one job: answering from the facts it's handed.
Measured, not promised
rikAI punches well above its weight — a significantly smaller model, significantly faster, with improved translation quality.
Scored blind by two neural judges and two string metrics on a frozen 660-sentence conversational set — the same bar every training run had to clear before it could ship.
One benchmark, every size
CyberAgent's CAT-Translate, untrained the same 0.8B, after rikAI's training its score, carried across the chart
Business Scene Dialogue — an external benchmark of everyday conversation, run English into Japanese. Higher is better on every panel. The grey line is the stock CAT-Translate family from CyberAgent, untrained; the gold square is the only model rikAI trains, so the gold stops at 0.8B. On all four metrics it lands above the stock 1.4B — a model nearly twice its size. Only the desktop-class 3.3B stays ahead.
Does size buy quality? Stock CAT-Translate on everyday Japanese
COMET-QE over 1,938 everyday sentences — Tatoeba gold (438) plus Tanaka conversational (1,500), plain prompts. Four times the parameters buys +0.003. On everyday Japanese, size stopped being the bottleneck — so rikAI ships the 0.8B and spends the savings on grounding.
What small buys — one sentence, translated
Greedy decoding, Apple silicon, measured at the shipping int8 precision. The 7B won't fit on a phone at all.
The reasoning model
A language model that explains grammar can't be allowed to improvise. RNF Reasoning was fine-tuned on exactly one skill: answer from the dictionary and grammar facts you're handed — and when the facts aren't there, say so. Before it shipped it had to pass seven grounding gates on a frozen 950-prompt eval. It passed all seven.
Sticks to the facts it's given counterfactual swap test
Admits it when the facts are missing grounded abstention
Invents synonyms the dictionary never said
stock 0.8B model RNF Reasoning, after fine-tuning
The gold-tagged inputs are computed, not generated — dictionary lookups and grammar matches are deterministic code, handed to both models as ground truth. The 0.8 billion parameters get spent only on what code can't do: writing the English, and talking to you.
RNF Translate 0.8B
the translator — its prompt, in full
→ writes the English, locked to those glosses
RNF Reasoning
the reasoning model — its briefing, in full
→ answers you, and can call the translator as a tool
In a small language model, anywhere you can be deterministic, you should be.
All of it, on the phone
Both models ship as int8 weights inside the app — the upgraded reasoning model is an optional ~0.8 GB download. Nothing you read, snap, or ask leaves the device. Airplane mode is a supported configuration.
Join the waitlist