AI News · industry
AI Translation Models in October 2026: Live Speech and Text
Google, OpenAI, Alibaba, DeepL, Bilibili and Tencent all shipped new translation models in 2026. What each one covers, where they still slip, and what it means if your customers write or call in other languages.
As of October 2026, AI translation models have split into two tracks. Live speech models from Google, OpenAI and Alibaba now listen in 60 to 70+ languages and talk back a few seconds behind the speaker, while open-weight text models such as Bilibili's Index-Translate cover 150 languages and can be downloaded for free. It matters for revenue: in a 2020 CSA Research survey of 8,709 consumers in 29 countries, 76% preferred to buy products with information in their own language, and 40% said they will never buy from websites in other languages.
The catch is in the fine print. "70+ languages" usually means 70+ languages the model can hear, not 70+ it can speak. Below is the global map, the gaps, and what we would check before trusting any of it with a customer.
What changed in AI translation between February and October 2026?
The table lists nine releases from labs in four countries: the US, China, France and Germany. The dates come from each vendor's own announcement.
| Date (2026) | Lab (country) | Release | Open or closed |
|---|---|---|---|
| Feb 12 | Kyutai (France) | Hibiki-Zero, simultaneous speech translation from French, Spanish, Portuguese and German into English | Open |
| Apr 16 | DeepL (Germany) | DeepL Voice-to-Voice, 40+ languages, API in early access | Closed |
| May 7 | OpenAI (US) | GPT-Realtime-Translate, 70+ input languages, 13 output languages | Closed |
| May 21 | Tencent (China) | Hy-MT2, 1.8B, 7B and 30B text models, 33 languages | Open |
| Jun 9 | Google (US) | Gemini 3.5 Live Translate, speech to speech, 70+ languages | Closed |
| Sep 18 | Alibaba (China) | Qwen3.8-LiveTranslate, 60 input languages, translated speech in 29 | API |
| Oct 1 | Microsoft (US) | MAI-Transcribe-2-Streaming and MAI-Voice-2.1, streaming transcription in 60 languages and one voice across 23 | Closed |
| Oct 2 | Bilibili (China) | Index-Translate-35B-A3B preview, text and image translation across 150 languages, Apache 2.0, with sibling Index-Echo speech-to-speech models | Open |
| Oct 6 | Mistral (France) | Mistral Large 4 preview, a general model (not a dedicated translator) trained on data spanning 160+ languages, weights promised by month end | Open (pending) |
India has its own track too. Sarvam-Translate, released in June 2025 and still Sarvam's open translation model, is built with AI4Bharat and covers all 22 scheduled Indian languages plus English, including structured long-form documents for 15 of them.
Which models translate live speech, and how many languages do they really speak?
Here is the number that most headlines skip. Input coverage is wide; output coverage is narrow.
| Model | Hears | Speaks translated audio in | Lag reported by vendor |
|---|---|---|---|
| Google Gemini 3.5 Live Translate | 70+ languages, detected automatically | 70+ languages (supported list) | "a few seconds" (Slator) |
| OpenAI GPT-Realtime-Translate | 70+ languages | 13 languages (OpenAI Cookbook) | keeps pace with the speaker |
| Alibaba Qwen3.8-LiveTranslate | 60 languages | 29 languages | 2.3 seconds average, down from 2.8 |
| DeepL Voice-to-Voice | 40+ languages | 40+ languages (API in early access) | not published |
| Kyutai Hibiki-Zero (open) | 4 languages | English | simultaneous |
OpenAI's 13 output languages are English, Spanish, Portuguese, French, German, Italian, Russian, Chinese, Japanese, Korean, Hindi, Indonesian and Vietnamese. Arabic and Turkish speakers can be understood by that model, but they cannot be answered in their own language with it. If your callers speak Arabic, that single row decides the question for you.
The newer systems also try to keep the caller's own voice. Google says its model preserves "intonation, pacing and pitch." Alibaba added speaker separation in September, so a three-way call no longer blends everyone into one translated voice. DeepL shipped new Voice models in September that keep each speaker's own voice, tone and rhythm across languages. We covered the interruption side of these voice models in our post on full-duplex voice AI.
Are dedicated translation models now beating general chatbots on text?
Sometimes, and it depends on which test you read.
Bilibili's own model card is unusually candid here. On FLORES (COMET-22), its 35B preview scores 0.8794, ahead of OpenAI's GPT-5.6 Sol at 0.8650 and Google's Gemini 3.5 Flash Lite at 0.8750 (a mid-tier Google model; Google's frontier is now Gemini 4 Argon). On the WMT26 judge score in the same table, though, GPT-5.6 Sol comes first with 89.10 against Index-Translate's 76.76. One caveat: GPT-5.6 Sol is also the model doing the judging in that test, and OpenAI has since moved ChatGPT to GPT-6. Neither number is a neutral leaderboard.
Tencent makes a similar claim for Hy-MT2: its paper reports that the 7B and 30B models beat DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, and that the 1.8B version fits in 440 MB after compression. That is small enough to run on a phone.
So, as of October 2026, there is no single "best translation model." Our read of the evidence:
- Specialist open models (Index-Translate, Hy-MT2, Sarvam-Translate) are now competitive with frontier chatbots on standard text benchmarks, at a fraction of the size. We wrote about why that pattern keeps repeating in small language models in 2026.
- Frontier general models still tend to win on judged, real-world fluency, and they handle whole conversations better than single sentences.
- Rare languages are where the gap opens. In Bilibili's low-resource table, Hy-MT2-1.8B answered in the wrong language 54.2% of the time, against 2.4% for Index-Translate-35B. That is one vendor scoring a rival, so treat it as a lead to test, not a verdict.
Where do these models still fall short?
Google's own documentation is the most useful source here, because it lists the weak spots plainly. Per Google's Live Translate limitations and Slator's summary of the model card, voice replication "may become inconsistent during long conversations or multi-speaker sessions," and language detection "can struggle with strong accents, similar languages, or rapid language switching."
Google adds that these detection slips mainly affect the input transcript, and that the final translation should still be accurate. Fair enough, but callers switch languages mid-sentence all the time, and a transcript your team can't trust is its own problem. Typed chat is a different case again: customers there mix scripts and dialects, writing Arabic in Latin letters or dropping English brand names into a Turkish sentence (our post on Arabic AI customer service in the Gulf has the dialect numbers). Live Translate takes audio only, so it does not solve that one.
Two more limits worth knowing:
- Names and product terms. Vendors are patching this, which tells you it is a known failure: DeepL added "customization with spoken terms," Qwen keeps hotword support for brands and product names, and Index-Translate accepts a glossary.
- A 2 to 3 second delay is fine for a lecture, awkward for a sale. Alibaba's 2.3-second average lag is good engineering. On a call where the customer is asking "so can I book Thursday?", it still feels like a satellite link.
What should a business check before relying on AI translation?
If you sell to customers who write or call in other languages, these are the questions that matter more than any benchmark:
- Which languages can it answer in? Ask for the output list by name; the input list is always longer.
- Does it translate, or does it reply natively? A model that writes directly in Spanish usually sounds more natural than English text passed through a translator.
- Can your team read every conversation? Someone has to review what the AI said, in a language they understand.
- Can you lock brand names and prices? A translated price or a mistranslated product name is a support ticket.
- What happens on a language it doesn't support? It should fall back to a language you chose, not guess.
How does Entagl handle conversations in other languages?
Entagl's AI Receptionist doesn't bolt a translator onto an English bot. It detects the customer's language on every message and writes the reply directly in that language, across 100+ languages, and it follows the customer when they switch mid-conversation. A one-word "ok" or a brand name doesn't flip the language back.
For your team, the inbox can translate a whole conversation on demand into any of its 10 team languages (English, Turkish, Arabic, Russian, Spanish, French, German, Italian, Portuguese and Greek), including voice-note transcriptions. When you take over and write a reply yourself, an AI rewrite turns your message into the customer's language before you send it. And if you only want the AI to answer in the languages you serve, an optional "Only reply in these languages" setting keeps every reply inside your list and falls back to your primary language for everything else. We walked through how those pieces fit together in multilingual customer service with AI and your team.
Entagl's product interface itself comes in 10 languages; the 100+ figure is what the AI agent answers customers in.
FAQ
What is the best AI translation model in 2026?
There isn't one winner. As of October 2026, open specialist models such as Bilibili's Index-Translate lead some text benchmarks (0.8794 on FLORES COMET-22 in its own report), while frontier general models still score higher on judged fluency. For live speech, of the models compared here, Google's Gemini 3.5 Live Translate has the widest spoken output at 70+ languages (supported list).
Can AI translate a phone call in real time?
Yes, with a delay. Alibaba reports a 2.3-second average lag for Qwen3.8-LiveTranslate, and Google describes its output as following "a few seconds behind." That works for meetings and support calls; it is noticeable in fast back-and-forth sales calls.
Are there free, open-source AI translation models?
Yes. Bilibili's Index-Translate (150 languages, Apache 2.0), Tencent's Hy-MT2 (33 languages), Sarvam-Translate (22 Indian languages) and Kyutai's Hibiki-Zero (speech, four languages into English) all publish their weights. Mistral says Large 4 weights will follow by the end of October 2026.
Do customers really care if a business answers in their language?
The best-known data says yes. CSA Research's survey of 8,709 consumers across 29 countries found 76% prefer buying with information in their own language and 40% will not buy from sites in other languages.
Should a small business use a translation model or a multilingual AI agent?
If customers message you on WhatsApp, Instagram or web chat, a multilingual agent that replies natively and lets your team read a translated view is simpler than wiring a translator into your inbox. Translation models make more sense for documents, subtitles and live events.
Customers writing in languages your team doesn't speak? See how Entagl answers, books and hands over in the customer's own language. Book a 30-minute demo.
Sources (checked October 21, 2026): vendor announcements and model cards from Google, OpenAI, Alibaba (via Slator), DeepL, Microsoft, Tencent (arXiv 2605.22064), Bilibili (Hugging Face), Mistral, Sarvam and Kyutai; consumer data from CSA Research (2020). Benchmark figures are as reported by each vendor and were not independently reproduced by Entagl.