The First Light · የመጀመሪያው ብርሃን

Ahadu

Native Amharic Intelligence

The first large language model built from the ground up for Ge’ez script. While other models translate Amharic through English, Ahadu thinks in it.

Read the paperUse the API

Model Family

Built for Ge’ez.
Engineered for the world.

Most large language models inherit their tokenizer from English. Ahadu’s 32K-vocabulary BPE tokenizer was trained from scratch on 50B+ Amharic tokens, representing Ge’ez script with 2–4× fewer tokens per sentence. Fewer tokens means lower cost, faster inference, and reasoning that follows Amharic structure rather than translating it.

Mixture-of-Experts · 1T total parameters · 32B active per pass · 256K context

Ahadu-Core-7B

7B · MoE active

General reasoning, chat, cultural knowledge.

Ahadu-Core-70B

70B · MoE active

Enterprise reasoning, complex analysis.

Ahadu-Bridge

3B · translation hub

Amharic ↔ 100+ languages.

Ahadu-Sino-7B

7B · zh ↔ am

Chinese–Amharic business and technical communication.

Ahadu-Lex-7B

7B · legal

Ethiopian legal code, case law, contract analysis.

Ahadu-Med-7B

7B · clinical

Diagnosis support, medical literature.

Ahadu-Code-1.5B

1.5B · code

Software engineering, code generation.

Ahadu-Vision-7B

7B · multimodal

Image understanding with Amharic description.

Live Demo

Ask in your language.

AhaduDemo mode · illustrative responses
ሰላም። I am Ahadu. Ask me about Ethiopian law, write in Amharic, or bring me a problem in code.

Technical Transparency

Weights, data, and scores — all public.

Open source

Apache 2.0. Model weights, the Ge’ez tokenizer, and the full evaluation harness are public from day one. Training used licensed and public-domain corpora only — no unauthorized distillation from proprietary models.

  • All 7 variants on open registries
  • Reproducible eval harness
  • Data mixture published per release

Data provenance

50B+ Amharic tokens, curated from Ethiopian legal code and case law, academic papers, literature, religious texts, news, and conversational data. Every dataset is licensed or public-domain, with the full mixture disclosed in the release card.

  • 50B+ Amharic tokens
  • Ethiopian legal corpus, curated
  • Explicit license per source

Benchmarks

Internal v1.0 evaluation against published baselines. Monochrome by design — the length of the bar is the only claim.

Amharic MMLU (higher is better)

78.2
61.4

Ahadu-Core-70B / GPT-4o baseline

Tokens per Amharic sentence (lower is better)

100
241

Ahadu Ge’ez BPE / English-first baseline · 2.4× fewer

Context window (tokens)

256K
128K

Ahadu-Core standard / typical frontier baseline

Ecosystem & API

One API. Every vertical.

# pip install ahadu
from ahadu import Ahadu

client = Ahadu(api_key=env.AHADU_API_KEY)

res = await client.chat.completions.create(
    model="ahadu-core-70b",
    messages=[
        {"role": "user", "content": "ስላም! ለምንዘጋጅባለህ?"},
    ],
    max_tokens=512,
    temperature=0.4,
)

print(res.choices[0].message.content)
# → ሰላም! ለአዲስ ስራ እዘጋጅባለሁ።
CommunityFree · 1K req/day · non-commercial
Builder$20/mo · 100K req · standard latency
Pro$200/mo · 1M req · priority latency
EnterpriseCustom · unlimited · SLA + dedicated support
On-PremiseLicense · full weights · air-gapped