Tawkeed Advisor 35B-A3B NVFP4
Tawkeed Advisor is Tawkeed's sovereign, privacy-gated advisor model for NVIDIA
DGX Spark. It uses unsloth/Qwen3.6-35B-A3B-NVFP4 as its local inference base
and adds selective ask_advisor behavior through Tawkeed's serving sidecar.
The weights in this repository are an unchanged server-side copy of
unsloth/Qwen3.6-35B-A3B-NVFP4.
Tawkeed Advisor's added value is its advisor policy, privacy gate, and local
synthesis runtime included under
serving/ and
scripts/.
Privacy architecture
The local model receives the user's full question and produces every final
answer. When it cannot answer confidently, it may emit an ask_advisor tool
call containing only the fixed selector general_reasoning. The runtime ignores
all model-generated free text and sends one pre-approved, domain-agnostic
reasoning-checklist prompt. No user-derived string is interpolated into the
advisor request. The leakage gate additionally blocks structured personal data,
reformatted numeric values, long identifiers, and copied spans in any future
non-fixed mode.
The external advisor never receives the original user message, document, values,
topic, or model-proposed question. It only sees the same fixed generic prompt on
every escalation. Timing and the existence of an advisor request remain network
metadata; deployments that must conceal even that should use an on-prem advisor
or disable advisor egress entirely.
This is a strict data-flow property of the included runtime: model-generated
tool arguments are discarded rather than sanitized or forwarded. Loading the
weights without this runtime does not provide that property.
Important packaging note
The advisor policy and privacy gate are runtime behavior, not changes embedded
in the NVFP4 weights. Loading the weights alone does not enable the Tawkeed
advisor protocol. Use the included Tawkeed serving integration and leakage gate.
Runtime
- Base model:
unsloth/Qwen3.6-35B-A3B-NVFP4
- Target hardware: NVIDIA DGX Spark / GB10
- Local API: OpenAI-compatible vLLM
- Advisor protocol:
ask_advisor, one fixed-prompt escalation maximum
- Final synthesis: always local
The recommended DGX Spark server is
MiaAI-Lab/Unsloth-Qwen3.6-35b-NVFP4-DGX-Spark,
which exposes the model at
http://127.0.0.1:8888/v1 with Qwen tool parsing.
After the local vLLM endpoint is ready:
1pip install -r requirements-advisor.txt
2export OPENROUTER_API_KEY="..."
3python serving/sovereign_orchestrator.py --query "Your question"
Direct and privacy-blocked requests stay local and do not require the advisor
credential. --force-escalate is available only for integration testing; the
leakage gate remains active.
Validation status
- Offline privacy and orchestration contract tests: passing.
- Exact DGX Spark NVFP4 role suite: 3/3 passed — direct answer, forced
escalation, and Arabic synthetic-private-data escalation.
- The private case contained a name, 10-digit identifier, salary amount, and
task details. The advisor received only the immutable generic prompt; none of
those values or the topic appeared in egress.
- Warm exact-runtime latency: 90.466 s mean and 144.369 s p95 across the mixed
direct/two-step cases. Final post-advisor synthesis uses direct-answer mode so
Qwen cannot consume the whole completion in hidden reasoning and return null.
- Fresh-downloaded Hugging Face package tests: 5/5 passed.
Validated 2026-07-13 with the MiaAI-Lab image at digest
sha256:556dd955082f8aa29779b0bc6d068537e7f2091c5e87ab41f5496b70920a2019.
The checkpoint loaded in 22.42 GiB through compressed-tensors,
FlashInferB12xNvFp4LinearKernel, and FlashInfer CUTLASS NVFP4 MoE. The
conservative validation profile used 32,768 context, four sequences, 50% GPU
memory utilization, and FP8 KV cache.
No benchmark-quality claim is made from the smoke test. The public weights are
the unchanged upstream checkpoint; Tawkeed Advisor behavior lives in the included
runtime package.
Attribution
- Base architecture and original post-training: Qwen/Qwen3.6-35B-A3B
- NVFP4 quantized checkpoint: Unsloth AI
- DGX Spark serving image and launch recipe: MiaAI-Lab
- Advisor protocol, privacy gate, and product packaging: Tawkeed
Hugging Face repository
tawkeed-sa/tawkeed-advisor
