Tawkeed Router — موجّه توكيد
Tawkeed Router is the local routing role in the Tawkeed Qwen3.6 family. It
selects exactly one downstream family without answering the user's request:
tawkeed-gpt for routine conversation, drafting, translation, extraction,
summarization, and straightforward questions.tawkeed-advisor for genuinely difficult specialist, mathematical,
scientific, or multi-step reasoning tasks where generic expert methodology
may improve the locally produced answer.
The weights are an unchanged server-side copy of
unsloth/Qwen3.6-35B-A3B-NVFP4.
The routing behavior is a runtime policy in
serving/tawkeed_router.py, not a
fine-tune embedded in the checkpoint.
Privacy and data flow
The default runtime is on-prem and makes no advisor or internet call. Its output
is strict JSON containing target, confidence, and a short category-level
reason. Invalid model output fails safely to Tawkeed Advisor. Sensitive data
alone does not trigger Advisor; routing is based on difficulty.
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:
1pip install -r requirements-tawkeed.txt
2python serving/tawkeed_router.py --query "ترجم هذه الجملة"
Validation status
- Offline routing contracts: passing.
- Exact DGX Spark NVFP4 role suite: 8/8 routing decisions passed.
- Warm exact-runtime latency: 5.561 s mean, 8.254 s p95 (sequential,
conservative eager-mode validation).
- Fresh-downloaded Hugging Face package tests: 2/2 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 compact
Qwen3-0.6B/head router is preserved separately as a legacy artifact and is not
the model represented by this repository.
