An open model that speaks the language of freight.
Inbound messages, documents, exceptions and operator commands, in English, Arabic, Urdu and Hindi, turned into structured operations data. Fine-tuned from Qwen2.5-7B, published with its benchmark, and small enough to run on-premise with no internet.
Try it on a message
Paste an email, WhatsApp or SMS the way your customers and drivers actually write them.
{
"origin": "Al Aweer Fruit & Vegetable Market",
"destination": "Mall of the Emirates",
"cargo_type": "chilled_produce",
"quantity": "26 pallets",
"weight_kg": null,
"vehicle_class": "reefer_4axle",
"pickup_date": "2026-10-16",
"pickup_time": "06:00",
"delivery_deadline": "2026-10-16 09:00",
"temperature_profile": "chilled",
"customer_ref": "BF-2291",
"missing": [
"weight_kg"
]
}Compliance, routing and cold-chain decisions are made by the Pangents engine, not by the model. The model reads, extracts, classifies and writes.
GulfFreightBench
660 held-out examples across six tasks plus a small real-email slice in English, Arabic, Urdu, Hindi and code-switched text. Field-level exact match; tool calls need the exact sequence and arguments. The written-message task is judged separately.
| Model | Overall | Triage | Intake | Docs | Exceptions | Tools | Valid JSON |
|---|---|---|---|---|---|---|---|
| Pangents Logistics LM 7B v0.2 | 0.943 | 0.958 | 0.844 | 1.000 | 0.994 | 0.920 | 100% |
| Qwen2.5-7B-Instruct (base) | 0.054 | 0.000 | 0.219 | 0.030 | 0.000 | 0.020 | 98–100% |
v0.2, 13 October 2026, 7,006 training rows, one NVIDIA L4, about 100 minutes. Benchmark v0.2 adds cross-border, unknown-cargo, imperial-unit and customs cases. The base model returns plausible JSON with invented field names, and in one case a fabricated API key inside a tool call.
The model never decides compliance.
Whether a truck may use a road at a given hour is a rule lookup in the Pangents engine. The model calls the engine; it does not replace it.
Labels correct by construction.
v0.1 training data is rendered from structured scenarios, so every label matches its text. Real anonymised operator messages are added in v0.2.
Honest benchmark.
GulfFreightBench uses places, customers and cargo vocabulary never seen in training. It still shares the generator’s style, so scores overstate raw-inbox accuracy. We say so.
Runs where your data is.
4-bit weights need about 5 GB. It runs on a laptop, an on-premise server, or a single cloud GPU, with no data leaving your environment.
Run it on your own messages
A 30-minute session: we run the model on a sample of your real inbox and show you what it extracts.