DCP
Open Navigation

Docs

ENAR

ALLaM 7B Instruct

Arabic-first Saudi model guide for deployment on DCP.

1. What it is

ALLaM 7B Instruct is a 7B-parameter Arabic-first instruction model from SDAIA/HUMAIN (model repo: `ALLaM-AI/ALLaM-7B-Instruct-preview`).

2. What it does

It is optimized for Arabic enterprise text workloads: assistant responses, summarization, and domain QA where Arabic quality is the top priority.

3. How it compares

  • Versus Falcon H1 7B: similar class, usually choose ALLaM when Arabic nuance matters more than raw throughput.
  • Versus JAIS 13B: JAIS can score higher on difficult Arabic tasks but costs more GPU memory and latency.

4. Best for on DCP

  • Arabic customer support copilots
  • Saudi public-sector and enterprise assistants
  • Arabic document understanding pipelines

5. Hardware requirements on DCP

  • DCP floor: `min_vram_gb: 24` (Tier A in `infra/config/arabic-portfolio.json`)
  • Recommended providers: RTX 4090, A100 40GB+
  • Template: `vllm-serve` for OpenAI-compatible API serving

6. How to run on DCP

  1. Submit a `vllm_serve` job with model `ALLaM-AI/ALLaM-7B-Instruct-preview`.
  2. Set `gpu_requirements.min_vram_gb` to `24`.
  3. Poll job output and endpoint readiness via `/api/jobs/:id/output` and `/api/jobs/:id/endpoint-ready`.

7. Licensing and commercial-use notes

ALLaM uses publisher-specific terms (not Apache/MIT). Confirm redistribution and commercial deployment rights directly from the model card terms before production launch.

Sources:

  • https://huggingface.co/humain-ai/ALLaM-7B-Instruct-preview
  • /home/node/dc1-platform/infra/config/arabic-portfolio.json
  • /home/node/dc1-platform/backend/src/db.js