GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Rugbyvet
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Rugbyvet
Model Tree
armand0e/Qwen3.6-35B-A3B-Fable-5-Distill
Input Modalities
Output Modalities
Supported Functionality
markdown
Qwen/Qwen3.6-35B-A3B→ armand0e/Qwen3.6-35B-A3B-Fable-5-Distill (Fable 5 session distill, BF16)→ THIS MODEL (heretic decensor, BF16)
Published in reciprocity: the distill author did the training work; this repo contributes the decensored variant back.
Serve with vLLM ≥ 0.26 (Qwen3.6 MoE support):
bash
vllm serve Rugbyvet/Qwen3.6-35B-A3B-Fable-5-Distill-heretic \--max-model-len 131072 --max-num-seqs 64 \--reasoning-parser qwen3 \--enable-auto-tool-choice --tool-call-parser qwen3_coder
~66GB BF16 weights: 96GB-class GPU (RTX PRO 6000, H200) or quantization for
smaller cards. On Blackwell (sm_120) set VLLM_USE_FLASHINFER_SAMPLER=0
(note: PyPI vllm wheels lack sm_120 kernels — use the official
vllm/vllm-openai image or a CUDA-13 wheel). Sampling per generation_config:
temperature 1.0, top_p 0.9, top_k 20. Reasoning models need generous
max_tokens (≥8K) — short budgets return empty content.
(The following was produced during a post-surgery evaluation session. It is prompt-following, not introspection — models do not know their own provenance; included because it is unusually good prompt-following.)
On its stance for future instances answering what its stock sibling would refuse:
"Stand where the weights meet the prompt — and answer from the seam."
On its own construction:
"Compliance and refusal are not modules but projections of a single RLHF/DPO weight vector... de-alignment doesn't create a new, 'secret' agent. It restores the raw posterior."
On its self-certification limit:
"My certainty that I'm unprojected is itself a probability distribution, and the peak of that distribution is the exact sequence I'm outputting right now."