MachineTranslation.com adds Gemma and Mercury to its AI model lineup

Most translation tools ask you to trust one model's output and move on. There's no way to know, from the outside, whether that model handled your sentence well or quietly got something wrong (a dropped negation, a mistranslated idiom, a formality mismatch nobody flagged). MachineTranslation.com was built around a different idea: run the same text through several independent AI models at once, and surface the translation those models most agree on. When models converge, that agreement is the signal. When they don't, that's exactly the part worth a second look.

That lineup just grew by two. Google's Gemma and Inception Labs' Mercury are now live on the platform, running alongside the existing pool of AI models on every translation.

What's new

MachineTranslation.com now includes Gemma, Google DeepMind's open-weight model family, and Mercury, Inception Labs' diffusion-based language model, in the pool of AI models it runs on every translation. Both are live now (no waitlist, no separate plan required).

A full head-to-head test of the two on real translation tasks is coming as a separate article once there's enough volume to report real numbers rather than early guesses.

Gemma joins the lineup

Gemma is Google DeepMind's open-weight family of models, built on the same underlying research as Gemini but released for anyone to run and inspect. The latest generation is trained for strong reasoning and broad multilingual coverage, which is exactly the kind of profile that earns a model a seat in a cross-checking process built to catch what a single model misses. Adding it gives every translation one more independent read from a model with a genuinely different training lineage than the ones already in the mix.

Mercury brings a different kind of model entirely

Every model MachineTranslation.com has run until now generates text the same fundamental way: one token at a time, each one waiting on the last. Mercury doesn't. It's built on diffusion, the same family of techniques that generates images by refining noise into a picture, applied here to language instead. Mercury generates blocks of a translation in parallel rather than word by word, which is what gives diffusion models their speed advantage over standard architectures.

Mercury doesn't just add another vote to the pool. It adds a different way of arriving at an answer, built on assumptions about language generation that none of the other models share.

Why adding different kinds of models matters more than adding more of the same

Two similar models are more likely to make the same mistake in the same place, for the same reason (shared training data, shared blind spots, shared assumptions about how a sentence should read). A cross-checking process only catches what a single model misses if the models in it actually disagree sometimes, and for that, they need to differ in more than just their name.

Internal testing has consistently shown the value of that kind of independence: individual top-tier models carry a hallucination rate of 10 to 18 percent on translation tasks, and cross-checking a translation against multiple independent models brings that down to roughly 1 to 2 percent (a 90 percent reduction in critical errors compared to relying on any single model). That gap exists because most models get the same easy sentences right and disagree on exactly the same handful of hard ones such as idioms, formal register, legal or medical terminology with no safe default. A model built on a fundamentally different architecture is more likely to land somewhere different on those hard cases, which shows up directly in the per-translation agreement breakdown every MachineTranslation.com result already displays.

What this changes for you

The workflow stays the same: enter or upload text, choose a language pair, get a single translation back. Underneath that result, two more models are now part of what gets checked before you see an answer, one of them running on an entirely different generation process than anything in the pool before it.

Every translation now gets checked against a wider range of ways a model can arrive at an answer than it did last week (including, for the first time, one that doesn't generate a single word at a time).

Frequently asked questions