GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
eliottmeunier
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
eliottmeunier
Model Tree
Base
Qwen/Qwen3-32B
Adapter
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
Trained so that integral reasoning lives in the weights, not in a system prompt. On a 12-dilemma benchmark scored with the deVos/AQAL rubric, the fine-tuned model moves the default center of gravity from orange/green (~3.35) to teal / 2nd tier (~4.5), and beats both a short "use the integral approach" prompt and a long mental-models system prompt on the base model. The AQAL scaffolding ("center of gravity", "partial truths", shadow) appears in the <think> channel on 13/13 test items.
unsloth/Qwen3-32B-unsloth-bnb-4bit (Apache 2.0).train_on_responses_only (loss on the full assistant turn, including <think>).<think> / answer pairs (English), generated with Claude under the integral analyst spec, system prompt then erased (Orca-style infusion).python
from unsloth import FastLanguageModelmodel, tok = FastLanguageModel.from_pretrained("eliottmeunier/qwen3-32b-integral-lora", # base + adaptermax_seq_length=3072, load_in_4bit=True)FastLanguageModel.for_inference(model)
Generation (Qwen3 thinking settings): temperature=0.6, top_p=0.95, top_k=20. The model emits a <think>...</think> block then the answer.
Judge and target share the AQAL/deVos frame, so the benchmark measures conformity to that frame, not absolute truth. n=12 dilemmas. The adapter installs a specific worldview (Wilber / deVos), on purpose. Credit: Corey deVos (Integral Life), Ken Wilber, Clare Graves.