Compute partnership · August 2026
A 27B model that works in Tunisian Arabic: chat, translation, diacritisation, and writing in a style you hand it. It reads images and video as well. It was trained on text there is nowhere to buy.
أحنا نبنيو أول موديل يفهم التونسي كيما يحكيوه التوانسة، بالدارجة، بالفرنساوي، والكل مخلوط.“We're building the first model that understands Tunisian the way Tunisians actually speak it.”
The spec
What we can tell you today.
- Family
- Barca H1
- Parameters
- 27B
- Does
- Chat, translation, diacritisation, style imitationall of it in Derja
- Also reads
- Images and videonot written up here yet
- Does not do
- Audio. Transcription is Fennec's
- Decode
- About 111 tokens a secondsingle stream
- Quality benchmark
- not yet
- Licence
- not yet
- Status
- Shipped
This page used to lead with a speech benchmark. That work is real and it is still published, on Fennec, which is the model that does the hearing and now scores better than the number that was here. What is left above is what has been measured about Hannibal itself. The two blank rows stay blank until somebody runs them.
Why we win
The models are the fast part. The engine behind them was not.
One pipeline feeds the whole lineup: the text this model was trained on, and the audio behind the other two.
01 · Collect
Audio nobody is selling
Tunisian TV and creator channels, pulled apart by source separation, diarization and segmentation. Two thousand hours of native speech and 1.31 million segments, all of it ours, because there is nothing on the market to buy.
02 · Vote
Four labelers that have to agree
Two independent ASR lineages, a CTC mixer and a Kaldi system. A transcript only gets in when they converge. Then a Derja-native language model checks the text actually sounds Tunisian, and hard filters throw out anything Egyptian or Moroccan that slipped through.
03 · Anchor
Every gate calibrated on humans
1,700 human-verified clips set the thresholds and 98.4 fully verified hours anchor each run. The audio is the asset we keep forever. The labels are disposable, and every new model rewrites them better than the last one did.
2,000+ h
of Tunisian speech collected and processed
421M
tokens of validated Tunisian text, Arabic script, Tunizi and MSA
1.31M
segments through the pipeline so far
369k
clips being relabeled by the champion right now
Why now
The expensive part is already paid for.
The corpus is built and three models are serving on one API. The only input we do not have is hours. Six months from now this is obvious and expensive; today it is four weeks of one machine.
API access
Put it in your product.
POST /api/v1/chat/completions · "model": "hannibal"
Hannibal has shipped: no waitlist, no access review. Sign in and create a key, shown once and stored hashed, and the first request is a copy-paste from the docs. Free while the beta lasts, 60 requests a minute.
No key yet? You can still talk to it in the playground.
Straight answers
What you're about to ask.
Can Hannibal transcribe audio?
No. It reads and writes, and it can look at an image or a video, but it does not hear. Transcription is Fennec, on POST /api/v1/audio/transcriptions. Hannibal served that endpoint while Fennec was in training and a request naming it there is now refused, which is better than handing back another model's transcript under this one's name.
Where did the speech numbers on this page go?
To Fennec, which is the model they belong to. They also improved on the way: the figure this page used to lead with is superseded. The whole table is published there, including the columns we have not rerun.
What is measured about Hannibal itself?
Decode speed, at about 111 tokens a second on a single stream, which is what you need to decide whether to stream a reply or wait for it. There is no published quality benchmark for Derja chat, ours or anyone else's, and this page will not invent one. The spec table says so rather than leaving the row out.
What stops a big lab from doing this next quarter?
The data. There is no Tunisian corpus to license, at any price. We collected 2,000 hours of speech and 421M tokens of text, then built the four-voter labeling stack and the human calibration set that decide what is good enough to train on. A lab starting today starts at the beginning of the pipeline, not the beginning of the model.