GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
tawkeed-sa
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
tawkeed-sa
Model Tree
unsloth/Qwen3.6-35B-A3B-NVFP4
Input Modalities
Output Modalities
Supported Functionality
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.
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.
unsloth/Qwen3.6-35B-A3B-NVFP4ask_advisor, one fixed-prompt escalation maximumThe 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:
bash
pip install -r requirements-mustashar.txtexport OPENROUTER_API_KEY="..."python 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.
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; Mustashar behavior lives in the included runtime package.
tawkeed-sa/tawkeed-mustashar-35b-a3b-nvfp4