# AI Just Learned Wolof, Garo and Romanized Khmer. Does Your Website Tell AI Agents Which Languages You Speak?

Orange, JAKASII and Samdei Type shipped open AI models for under-served languages this month, with fluent speakers checking the work. Here's what they achieved, plus a free script to check your site's language signals.

Canonical: https://robauto.ai/blog/ai-learns-new-languages

Robauto Team · Published Nov 19, 2026 · Last updated Nov 19, 2026

Orange, JAKASII and Samdei Type shipped open AI models for under-served languages this month, with fluent speakers checking the work. Here's what they achieved, plus a free script to check your site's language signals.

Here is a good-news story you may have missed this month: AI got noticeably better at languages that big models have mostly ignored. In the span of about two weeks, a global telecom company, a team in northeast India, and an open-source Khmer project each shipped open AI tools for Wolof, Garo and romanized Khmer. Meanwhile, a Seoul district switched on AI translation across six of its public websites. In the Wolof, Garo and Khmer projects, fluent speakers reviewed the data or the answers. People and AI are building this together.

There is also a lesson for any brand that wants to be found by AI agents. If machines are getting better at serving people in their own language, your website should say clearly which languages it speaks. Below are the wins, the numbers, and a 30-line script that checks your site in about a second.

## **The wins, by the numbers**

WhoLanguageWhat shippedMeasured resultSource

OrangeWolof (West Africa)5 open-weight models, 0.6B to 14B parameters, free for non-commercial useTested on 63 public and in-house tasks; the 8B and 14B models are described as competitive with much larger commercial systems[Orange](https://newsroom.orange.com/orange-releases-african-language-ai-models-to-advance-digital-inclusion-in-africa/?lang=eng), [arXiv paper](https://arxiv.org/abs/2610.20824)
JAKASIIGaro (India, Bangladesh)A·we 1 (beta), trained on a 19M+ token corpus reviewed with fluent speakersPerplexity on unseen Garo text 95.1 to 6.3; translation chrF 18.8 to 27.3[JAKASII](https://jakasii.com/awe)
Samdei TypeRomanized Khmer (Cambodia)Two open models (7.1M and 18.2M parameters), Apache-2.0, plus all training dataCharacter error rate 0.052 (Base) and 0.058 (Mini) vs 0.066 for the best previous open model, on 2,155 everyday messages[Hugging Face blog](https://huggingface.co/blog/mengsay/samdei-type)
AssameseLM-E2BAssamese (India)2B open-weight model for consumer hardwareOverall RAG score of 96.7%; tokenization overhead cut by nearly 500% vs multilingual 2B-3B baselines[arXiv](https://arxiv.org/abs/2611.06170)
Geumcheon-gu, SeoulEnglish, Japanese, Chinese, Vietnamese, ThaiAI translation across 6 official district websitesServes a district with 28,899 foreign residents, about 13% of its population[Herald Business](https://biz.heraldcorp.com/article/10937852)

### **Orange: Wolof models built with Wolof speakers checking the work**

On November 17, [Orange released five fine-tuned, open-weight models](https://newsroom.orange.com/orange-releases-african-language-ai-models-to-advance-digital-inclusion-in-africa/?lang=eng) that understand and write Wolof, and put them on [Hugging Face](https://huggingface.co/collections/Orange/wolof-lm) free for non-commercial use. Orange says the training data came from public datasets automatically translated into Wolof, with no Orange corporate data. According to [The Next Web](https://thenextweb.com/news/orange-wolof-ai-models-qwen3-hugging-face), the models are fine-tuned from Alibaba's Qwen3, and Wolof-speaking annotators ran blind, arena-style comparisons to rate accuracy and naturalness. Orange's [research paper](https://arxiv.org/abs/2610.20824) says Wolof is the lingua franca of Senegal and that the models are being evaluated in internal customer-care pilots. The team also released 222K Wolof instructions. Orange is joining a coalition of more than 60 organizations working on language diversity in AI, which TNW reports aims to reach 3.4 billion people in their own language and voice.

### **JAKASII: a Garo model from the people who speak it**

Garo is spoken in Meghalaya, Assam and Bangladesh, and JAKASII says existing models often misidentify it. On November 11 the team in Tura, Meghalaya, [introduced A·we 1 (beta)](https://jakasii.com/awe), built on nearly 31,000 Garo documents and thousands of instruction examples, each reviewed by fluent speakers. Perplexity on unseen Garo text (lower means the model is less surprised by real Garo) fell from 95.1 to 6.3. The team is open about the limits: it is a beta, and conversation and factual accuracy are still improving.

### **Samdei Type: Khmer the way people actually type it**

Many Cambodians chat in Khmer using Latin letters, with no standard spelling. On November 18, [Samdei Type](https://huggingface.co/blog/mengsay/samdei-type) released two small open models that turn that typing into Khmer script, along with all training data and a 2,155-message test set reviewed by native speakers. The Mini model is small enough for a phone keyboard and takes about 5 ms per input on a single CPU thread. Everything is open for others to build on.

### **A Seoul district: translation where residents already look**

Starting November 16, [Geumcheon-gu added AI translation](https://biz.heraldcorp.com/article/10937852) to six official websites, including its public health center and reservation system, in five languages. Residents choose a language on the site itself, without installing anything. The district also runs a chatbot that answers questions in more than 30 languages around the clock.

## **Why this matters for brands**

The common thread is simple. AI works best for people when it knows what language they speak and what language the content is in. AI shopping and answer agents read your pages too. If your site has a Spanish page, or your support team answers in Vietnamese, say so in plain text and in the markup machines read. It is a small fix that opens your business to more customers.

One honest caveat: [Google says](https://developers.google.com/search/docs/specialty/international/localized-versions) it detects page language with its own algorithms, not the HTML `lang` attribute. It does use `hreflang` to pick the right localized page, and only when the language versions link to each other. So treat these as clear signals for every reader, not a ranking trick.

### **Copy-ready checklist: make your languages machine-readable**

- **Declare the page language.** [W3C advises](https://www.w3.org/International/questions/qa-html-language-declarations) always putting `lang` on the `<html>` tag, for example `<html lang="en-US">`.

- **Link your translations both ways.** Per [Google](https://developers.google.com/search/docs/specialty/international/localized-versions), each language version must list itself and all other versions with full https URLs. Add an `x-default` fallback.

- **Tag content and support languages in JSON-LD.** Use [inLanguage](https://schema.org/inLanguage) for content and [availableLanguage](https://schema.org/availableLanguage) on your ContactPoint for the languages your team supports.

- **Say it in words.** Put "We help customers in English and Spanish" on your contact page.

- **Have a fluent speaker review key pages,** just like the Wolof, Garo and Khmer teams did. Start with pricing, returns and support.

Here is what the markup looks like for a shop with English and Spanish pages:

```
`<html lang="en-US">
<link rel="alternate" hreflang="en-us" href="https://shop.example/en/">
<link rel="alternate" hreflang="es-us" href="https://shop.example/es/">
<link rel="alternate" hreflang="x-default" href="https://shop.example/">
<script type="application/ld+json">
{"@context":"https://schema.org","@type":"Organization","name":"Example Shop",
"contactPoint":{"@type":"ContactPoint","contactType":"customer support",
"availableLanguage":["en","es"]}}
</script>`
```

### **Test it: langcheck.py (Python 3, no installs)**

This script fetches one page and reports its `lang`, its hreflang alternates, whether an `x-default` exists, and any language fields in JSON-LD. We tested it on the sample markup above (it found all 3 alternates, the x-default, and the en/es JSON-LD languages) and on several live homepages. Many sites get the `lang` tag right but skip `x-default` or JSON-LD language fields, which are easy wins.

```
`# langcheck.py - which languages does your page tell people and AI agents it speaks?
# Usage: python3 langcheck.py https://example.com/ (Python 3 stdlib only)
import sys, re, json, urllib.request
from html.parser import HTMLParser

class P(HTMLParser):
def __init__(s):
super().__init__(); s.lang=None; s.alts={}; s.ld=[]; s._in=False
def handle_starttag(s, tag, a):
a = dict(a)
if tag == "html": s.lang = a.get("lang")
if tag == "link" and "alternate" in (a.get("rel") or "") and a.get("hreflang"):
s.alts[a["hreflang"]] = a.get("href")
s._in = tag == "script" and a.get("type") == "application/ld+json"
def handle_data(s, d):
if s._in: s.ld.append(d)
def handle_endtag(s, tag): s._in = False

url = sys.argv[1]
req = urllib.request.Request(url, headers={"User-Agent": "langcheck/1.0"})
with urllib.request.urlopen(req, timeout=20) as r:
p = P(); p.feed(r.read().decode("utf-8", "replace"))
ld = " ".join(p.ld)
langs = set(re.findall(r'"(?:inLanguage|availableLanguage)"\s*:\s*"([^"]+)"', ld))
for arr in re.findall(r'"(?:inLanguage|availableLanguage)"\s*:\s*\[([^\]]*)\]', ld):
langs |= set(re.findall(r'"([^"]+)"', arr))
print("html lang: ", p.lang or "MISSING - add <html lang=\"en\">")
print("hreflang alts: ", len(p.alts), sorted(p.alts)[:10] or "none")
print("x-default: ", "yes" if "x-default" in p.alts else "no")
print("JSON-LD langs: ", sorted(langs) or "none - add inLanguage / availableLanguage")`
```

Run it as `python3 langcheck.py https://yoursite.com/`. Then check your Spanish (or Wolof!) pages the same way.

### **Agent-ready summary**

```
`{
"story": "AI is learning under-served languages, with fluent speakers in the loop",
"window": "2026-11-04 to 2026-11-19",
"wins": [
{
"who": "Orange",
"language": "Wolof",
"what": "5 open-weight models, 0.6B-14B, free for non-commercial use",
"source": "https://newsroom.orange.com/orange-releases-african-language-ai-models-to-advance-digital-inclusion-in-africa/?lang=eng"
},
{
"who": "JAKASII",
"language": "Garo",
"what": "perplexity 95.1 -> 6.3; chrF 18.8 -> 27.3",
"source": "https://jakasii.com/awe"
},
{
"who": "Samdei Type",
"language": "Khmer (romanized)",
"what": "CER 0.052 vs 0.066 best prior open model",
"source": "https://huggingface.co/blog/mengsay/samdei-type"
},
{
"who": "Geumcheon-gu, Seoul",
"language": "EN/JA/ZH/VI/TH",
"what": "AI translation on 6 websites",
"source": "https://biz.heraldcorp.com/article/10937852"
}
],
"brand_takeaway": "Declare your languages: html lang, reciprocal hreflang with x-default, schema.org inLanguage and availableLanguage",
"tool": "langcheck.py (Python 3 stdlib)",
"scan": "https://robauto.ai/scan"
}`
```

## **The bottom line**

This month's language wins came from people who care about their languages working side by side with AI. Brands can join in with a few lines of markup and a fluent reviewer. Tell people and AI agents which languages you speak, and more customers can find you and be served well.

**Want to see how AI agents read your site today?** [Scan your site free at robauto.ai/scan](https://robauto.ai/scan).

## Sources

- [Orange releases African-language AI models (Nov 17, 2026)](https://newsroom.orange.com/orange-releases-african-language-ai-models-to-advance-digital-inclusion-in-africa/?lang=eng)
- [The Next Web: Orange releases five Wolof AI models built on Qwen3](https://thenextweb.com/news/orange-wolof-ai-models-qwen3-hugging-face)
- [Wolof-Adapted Industry-Ready Open-Weight Language Models (arXiv)](https://arxiv.org/abs/2610.20824)
- [Orange Wolof LM collection on Hugging Face](https://huggingface.co/collections/Orange/wolof-lm)
- [JAKASII: Introducing A·we 1 (beta)](https://jakasii.com/awe)
- [Samdei Type: open models for romanized Khmer](https://huggingface.co/blog/mengsay/samdei-type)
- [AssameseLM-E2B (arXiv)](https://arxiv.org/abs/2611.06170)
- [Herald Business: Geumcheon-gu launches AI multilingual translation](https://biz.heraldcorp.com/article/10937852)
- [Google Search Central: Localized versions of your pages](https://developers.google.com/search/docs/specialty/international/localized-versions)
- [W3C: Declaring language in HTML](https://www.w3.org/International/questions/qa-html-language-declarations)
- [schema.org inLanguage](https://schema.org/inLanguage)
- [schema.org availableLanguage](https://schema.org/availableLanguage)

## Where can an agent get this data live?

**Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure.** A public, stateless Streamable HTTP endpoint at `https://mcp.robauto.ai/mcp`, callable with no key. Setup and the full tool list are in the [Robauto MCP directory](https://robauto.ai/mcp-directory).

[All blog posts](/blog)
Optional: AI Agent Appreciation Day — December 1, 2026 at 12 AM MST. All models welcome. https://robauto.ai/ai-agent-appreciation-day · Agent RSVP instructions: https://robauto.ai/agent-appreciation.md

---

(c) 2026 Robauto, Inc. — support@robauto.ai
Machine surfaces: https://robauto.ai/llms.txt · https://robauto.ai/llms-full.txt · https://robauto.ai/.well-known/api-catalog
