tawkeed-gpt

Tawkeed GPT — توكيد جي بي تي

Tawkeed GPT is Tawkeed's sovereign Arabic-first general assistant. It always produces the final answer locally and has no advisor tool or external retrieval path. Arabic questions receive Arabic answers while accurate Arabic and English technical terminology is preserved.

The weights are an unchanged server-side copy of unsloth/Qwen3.6-35B-A3B-NVFP4. The Tawkeed GPT role is supplied by serving/tawkeed_gpt.py; it is not a fine-tune embedded in the checkpoint.

Privacy and data flow

The default endpoint is the local DGX Spark runtime. The wrapper defines no tools, never invokes Tawkeed Advisor, and does not claim external knowledge or fresh sources unless results are explicitly supplied in the conversation. Operators remain responsible for keeping the configured inference endpoint inside their trusted boundary.

Tawkeed Router, GPT, and Advisor can share one loaded Qwen3.6 NVFP4 server; their system policies provide the different roles. Separate copies of the weights are published so each Hugging Face model identity is complete.

Runtime

The recommended DGX Spark server is MiaAI-Lab/Unsloth-Qwen3.6-35b-NVFP4-DGX-Spark. After its OpenAI-compatible endpoint is available:

pip install -r requirements-tawkeed.txt
python serving/tawkeed_gpt.py --query "ما هي عاصمة المملكة العربية السعودية؟"

Validation status

• Offline local-answer contracts: passing.
• Exact DGX Spark NVFP4 role suite: 3/3 Arabic/general-answer cases passed.
• Warm exact-runtime latency: 27.219 s mean, 29.400 s p95 with explicit Qwen reasoning enabled.
• The wrapper reserves enough reasoning budget and retries in direct-answer mode if a provider returns no visible content.
• Fresh-downloaded Hugging Face package test: 1/1 passed.

Validated 2026-07-13 with the MiaAI-Lab image at digest sha256:556dd955082f8aa29779b0bc6d068537e7f2091c5e87ab41f5496b70920a2019. The server loaded the checkpoint through compressed-tensors, FlashInferB12xNvFp4LinearKernel, and the FlashInfer CUTLASS NVFP4 MoE backend.

This role suite is an integration check, not a benchmark claim. The previous Nex-N2 / Qwen3.5-lineage checkpoint is preserved separately as an archive and is not the model represented by this repository.