Orange fine-tunes Qwen3 for Wolof: five open models from 0.6B to 14B parameters
Chatbots handle languages with little web text poorly, and Wolof, spoken by about 16 million people in West Africa, is one of them. Orange fine-tuned Alibaba's open Qwen3 model and released five versions from 0.6B to 14B parameters: the small ones run on weak hardware, while the large one is more accurate. For a Wolof support bot this is a ready-made foundation, but a paid product requires an agreement with Orange.
AI-processed from TNW; edited by Hamidun News
Imagine you are building a support chatbot for a bank or a telemedicine service in Senegal. Any popular model answers decently in French and English, but your customers speak Wolof, a language of about 16 million people in West Africa. There is little Wolof text on the internet, so standard models get confused or answer in stilted, word-for-word phrasing.
Telecom operator Orange has released five open models for exactly this case: it took the Qwen3 family from China's Alibaba and fine-tuned it on Wolof. The numbers in the model names are their sizes, from 0.6 to 14 billion parameters (parameters are a neural network's "settings"; the fewer there are, the easier the model is to run on weak hardware, and the more there are, the more capable it usually is).
For developers, this means Wolof finally has an open foundation for building prototypes.
What exactly was released
Orange presented the models at the OpenTech event in Paris. All five have open weights: you can download them and run them yourself. They are free for non-commercial use, while commercial use, according to the company, is possible under a separate agreement with Orange. So a startup that wants to embed a model in a paid service will first have to negotiate with the operator.
The range of sizes is practical: you can build a prototype on the lightest version with 0.6 billion parameters, then move up to the largest one with 14 billion when you need better quality.
The training data consists of public datasets that were automatically translated into Wolof. Orange's corporate data is not part of the training. That settles some questions about data sources, although machine translation inevitably leaves unnatural phrasing in the texts.
Why Qwen, and not Llama and Whisper
The release came a year after the announcement. In November 2024, Orange planned to fine-tune Meta's Llama text model and OpenAI's Whisper speech model, with Wolof and Pulaar due to appear in the first half of 2025. The base model was changed in the end. A company representative told TNW that Qwen was chosen because at the time it gave the best quality for language modeling. Whisper is a speech recognition model, not a language model, so it was not suitable for a text task.
Humans checked the quality. Native Wolof speakers evaluated both the training data and the models' answers. They ran Arena-style blind comparisons across different tasks and scored accuracy and naturalness of language separately. The approach makes sense: an answer can be factually correct but sound unnatural, and automatic metrics catch that poorly. Orange did not provide specific figures on the results, so you will have to compare the models with alternatives yourself.
What's next and what developers should do
Pulaar, announced together with Wolof, belongs to Orange's voice work rather than its text models, and remains in the plans. For now, voice support for other languages takes priority: Darija (Morocco), Dyula and Mossi (Burkina Faso).
Orange has also joined a language AI partnership launched in September. It has more than 60 participants, and its goal is to make AI accessible to 3.4 billion people in their native language and voice.
What to do with this in practice:
- If you are working on services for francophone Africa (support, banking, education, healthcare), download the models and test them on real customer queries in Wolof: there are no independent benchmarks yet.
- Orange's approach is reproducible for other small languages: a strong open base model, fine-tuning on translated public data, and evaluation by native speakers.
- Plan for licensing early: a paid product will require a separate agreement with Orange.
For the industry, this is one more example of open Chinese models like Qwen becoming the base for local language projects, including at European corporations.
- Nov 20, 2026Anthropic goes public: shares open at $111.50 versus an $83 offer price, up 34%
- Nov 20, 2026Google teaches Gemini to run your Mac in the background without taking over the screen or cursor
- Nov 20, 2026Yandex 360 Launches MCP Server, Giving AI Agents Access to 7 Office Services
- Nov 20, 2026Gemini API Cuts Token Use for a 200-Tool Agent Nearly 7x, by Google's Estimate
Key Questions Answered
What exactly was released?
Orange presented the models at the OpenTech event in Paris. All five have open weights: you can download them and run them yourself. They are free for non-commercial use, while commercial use, according to the company, is possible under a separate agreement with Orange. So a startup that wants to embed a model in a paid service will first have to negotiate with the operator.
Why Qwen, and not Llama and Whisper?
The release came a year after the announcement. In November 2024, Orange planned to fine-tune Meta's Llama text model and OpenAI's Whisper speech model, with Wolof and Pulaar due to appear in the first half of 2025. The base model was changed in the end. A company representative told TNW that Qwen was chosen because at the time it gave the best quality for language modeling. Whisper is a speech recognition model, not a language model, so it was not suitable for a text task. Humans checked the quality.
What's next and what developers should do?
Pulaar, announced together with Wolof, belongs to Orange's voice work rather than its text models, and remains in the plans. For now, voice support for other languages takes priority: Darija (Morocco), Dyula and Mossi (Burkina Faso). Orange has also joined a language AI partnership launched in September. It has more than 60 participants, and its goal is to make AI accessible to 3.4 billion people in their native language and voice.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.