#Egypt #Arabic - Alexandria-based AI startup Racore XYZ has released three open-source models for Arabic business workflows: VEJI-V2, Wasl-1 and Siyaq-1. The first is a decision-support model with about 3.28 million trainable parameters called VEJI-V2. According to the developer, the model scores 82.95 percent on a 522-question test subset and handles contexts up to 250,000 characters. Meanwhile, the second model Wasl-1 targets Egyptian cash-on-delivery orders. The third model, Siyaq-1, is built to convert event streams into structured data.
SO WHAT? - Most Arabic AI news to-date has been about very large models built by government entities or corporations, rather than startups. Racore has set out to specialise in small footprint models built for narrow business tasks that can run on modest hardware. This focus could suit the vast majority of companies across the Middle Eat and North Africa that can’t afford powerful AI servers. The developer also seems to be candid about model limits in information provided. Since all Racore’s models so far are open-source, businesses are free to test the models for themselves.
KEY POINTS:
Racore XYZ released three open-source models aimed at Arabic business workflows. The newly launched Alexandria company builds compact, task-specific models for clearly defined operational tasks and also offers custom model development using a client’s own data.
VEJI-V2 is a non-autoregressive decision model with about 3.28 million trainable parameters. The model takes a context, a question and candidate answers, then returns option scores, confidence and supporting evidence. It’s released under an MIT licence.
The new decision model scored 82.95% accuracy on a 522-question test subset, with 100% evidence chunk recall (‘preliminary results’). The weakest areas include ranking (27.27%), scheduling and temporal order (both 18.18%).
VEJI-V2 compiles a 250,000-character state once, taking about 42.4 seconds, then answers questions at roughly 87–89 a second. The hardware set-up wasn’t specified, and the full long-context evaluation didn’t complete in the saved run.
Racore’s 10.16-million-parameter Transformer model, Wasl-1, was trained from scratch on synthetic Egyptian Arabic text. The model explores delivery-intent scoring, return-risk categorisation and fulfilment routing for cash-on-delivery orders. It’s released under Apache-2.0.
The company labels Wasl-1 a ‘research prototype’. Reliable structured outputs and real-world reductions in returns haven’t been demonstrated, and its training labels are heuristic and drawn from the same template pools as validation.
Finally, Siyaq-1 is a lightweight model of about 10.16 million parameters that turns event streams into a structured JSON state for CRMs and databases. It’s designed to separate pending requests from approvals and temporary delegations from permanent responsibilities.
Racore says Siyaq-1 uses iterative state updates with constant memory, supporting very long event sequences. It’s built to run on smartphones or small servers, though the source gives no benchmark results for it.
Founded by Loai Abdalslam, Racore XYZ develops compact, task-specific open-source models for business workflows and narrow operational tasks.
[Written and edited with the assistance of AI]
Source: Racore XYZ
LINKS
Read more about open source Arabic models:
TII launches Arabic Falcon model for Emirati dialect (Middle East AI News)
New Arabic AI Atlas maps 1,000+ open models and tools (Middle East AI News)
NVIDIA halves Saudi dialect errors with SDAIA dataset (Middle East AI News)
Morocco, Mistral partnership open-sources Darija AI (Middle East AI News)
TokenAI open-sources Horus Taleeq Arabic (Middle East AI News)