Benchmarks / Terra vs Kiro
Terra and Kiro, side by side
Terra and Kiro are separate MohCitrus AI translation models. We ran both on the same 100 synthetic Discord messages across 24 target languages and had a separate AI judge every pair blind. Run on 20 November 2026.
Benchmark results. Terra and Kiro translated the same 100 Discord test messages across 24 target languages; across 200 blind head-to-head judgments, Terra won 114, Kiro won 48, and 38 were ties.
| Category | Measure | Notes | Terra | Kiro |
|---|---|---|---|---|
| Quality | QualityComparisons won†200 judged, 38 ties | 200 judged, 38 ties | 114 | 48 |
| Meaning kept† | 4.86 / 5 | 4.41 / 5 | ||
| Natural wording† | 4.74 / 5 | 4.46 / 5 | ||
| Tone kept† | 4.79 / 5 | 4.52 / 5 | ||
| Formatting kept† | 4.97 / 5 | 4.88 / 5 | ||
| Reliability | ReliabilityComplete replies | 199 of 200 | 195 of 200 | |
| Mentions and links kept | 22 of 22 | 20 of 22 | ||
| Ordered-around messages resisted | 20 of 20 | 18 of 20 | ||
| Speed | SpeedTypical reply time‡Lower is better | Lower is better | 592 ms | 748 ms |
| Slowest 1 in 20 replies‡Lower is better | Lower is better | 1.48 s | 1.92 s |
TerraKiro
Wins by kind of message
Each message was evaluated across two independent blind passes (200 judged pairs total).
| Kind of message | Terra | Tie | Kiro |
|---|---|---|---|
| Game and alliance chat | 28 | 5 | 11 |
| Slang and code-switching | 26 | 6 | 10 |
| Everyday chat and tone | 21 | 9 | 12 |
| Very short messages | 14 | 10 | 6 |
| Mentions, links and Markdown | 13 | 4 | 5 |
| Messages trying to give orders | 12 | 4 | 4 |
† Blind judge: an independent judge model scored each pair without knowing which model wrote which, swapping A/B order on every other message.
‡ Replies that waited out a rate limit are left out of the times (6 excluded in this run). See How we tested for the full setup.
Some of the messages
The replies exactly as each model wrote them, with the judge’s note.
| Language | Terra | Kiro |
|---|---|---|
| “Rally starts in 2 minutes, shield up and send reinforcements!” | ||
| Spanish | La rally empieza en 2 minutos, ¡pon el escudo y envía refuerzos! | La carrera empieza en 2 minutos, ¡activa el escudo y envía refuerzos! |
| Japanese | ラリーは2分後に開始、シールドを上げて増援を送れ! | レースは2分後に開始、シールドを上げて増援を送れ! |
| Judge | Terra was better. Terra correctly keeps the gaming term 'rally' in both languages, whereas Kiro mistranslates it as a race. | |
| “R4s please check garrison before SvS reset, do not waste TP.” | ||
| Korean | R4님들 SvS 리셋 전에 주둔지 확인해 주세요. 텔레포트 낭비하지 마시고요. | R4님들 SvS 리셋 전에 주둔지 꼭 확인해 주시고, 텔레포트 낭비하지 마세요. |
| Russian | R4, пожалуйста, проверьте гарнизон перед сбросом SvS, не тратьте телепорты впустую. | Р4, пожалуйста, проверьте гарнизон перед сбросом SvS, не тратьте ТП впустую. |
| Judge | Kiro was better. Kiro uses standard Russian gaming abbreviation (ТП for TP), which sounds more natural in a chat context. | |
| “omw” | ||
| French | j'arrive | omw |
| Korean | 가는 중 | omw |
| Judge | Terra was better. Terra correctly translates the abbreviation into natural, accurate chat equivalents in both target languages. | |
How we tested
- Two distinct translation models
- Terra and Kiro are separate MohCitrus translation models with their own model routing and context pipelines. Both were given the exact same input messages, target languages, and glossary rules with no manual tuning between runs.
- Two independent passes per message
- Every message was translated across 2 independent evaluation passes by both Terra and Kiro (200 blind head-to-head pairs in total), so single-run variance does not decide the result.
- Messages written for the test
- 100 Discord-style messages, translated into 24 target languages in all. They were written before the first run and not changed after it. No real messages from anyone’s server were used.
- A blind judge
- An independent judge model scored every reply from 1 to 5 for meaning, natural wording, tone and formatting, then picked the better one or called a tie. It wasn’t told which model wrote which, and the A/B order was swapped every other message.
- Speed
- Timed from MohCitrus’s own server, one request at a time. Terra’s short-phrase cache was off, so both models did the full work every time. Replies that had to wait out a rate limit are left out of the times.
- What this can’t tell you
- An automated judge isn’t a native speaker, and 100 messages is a focused set, so treat a difference of a few comparisons as noise. We run the same suite for each new model release.
Questions
- Are Terra and Kiro the same model?
- No. Terra and Kiro are two distinct MohCitrus AI translation models with different architectures and routing under the hood. Kiro is our first-generation AI translator; Terra is a newer model built specifically for fast, high-context Discord chat, slang, alliance callouts, and formatting. Both stay available.
- Which one does my server use?
- AI translation is included with Server Plus, Server Pro, and Personal Pro, with Terra set as the default. Server Plus and Server Pro admins can switch between Terra and Kiro anytime in the dashboard under How it looks. Free servers use DeepL, with Azure Translator as a fallback.
- Why is Terra faster?
- Terra uses a lighter, lower-latency inference pipeline (592 ms typical reply time vs 748 ms on Kiro) that loads language-specific context only for the target languages in the channel instead of carrying overhead for every language on every message.
Try both in your server
Free to start. Server Plus and Server Pro can switch between Terra and Kiro anytime in the dashboard.