GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
U-WIN
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
U-WIN
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
On the il protocol (the fixed 100 Stage II held-out questions, 50 TRUE and 50 FALSE, 16 attempts each, Qwen3.8-27B judge with the reference rubrics, six grading passes): mean rubric score 0.409 (best of 16: 0.662; questions with a perfect attempt: 54 of 100; TRUE questions 0.505, FALSE questions 0.313; mean completion 9,819 tokens, 194 of 1,600 attempts hit the 24,576-token cap). Reference points on the same protocol: base Qwen3.5-9B 0.330, notes-only 50M + KL 0.392. Paired difference against the notes-only checkpoint over the same questions: +0.017 (95% bootstrap interval +0.003 to +0.033). These numbers are not on the scale of the canonical Stage II table (205 questions, gpt-5.6 judge, 4 attempts) kept in U-WIN/equational-theory-thinking-7030-evals; this checkpoint has not been scored there.
The full report (every dataset with examples, all mixes, the paired comparisons and a behavioral analysis of the traces) is the
private U-WIN Space equational-internalization-results. Training metrics: W&B run.
The safetensors are the exact serving export that was evaluated (427 fine-tuned language tensors plus 348 unchanged base tensors; text weights were fine-tuned, the vision components are the base model's and were not evaluated for multimodal quality). The complete config, tokenizer, chat template and processors are included, so no local graft or adapter loading is
needed. training_summary.json carries the provenance (launch record digests, checkpoint, tensor audit) and artifact_manifest.json
the sha256 of every file.
Model downloads require an individual request and manual approval by U-WIN. U-WIN organization members retain the organization's built-in access; applying the gate to non-admin members requires a paid Hugging Face organization plan. Admins and the repository creator always retain direct access. Access requests are reviewed in this repository's Settings page.
Use a Qwen3.5-compatible Transformers installation and an approved HF token:
python
from transformers import AutoModelForImageTextToText, AutoProcessor model_id = "U-WIN/qwen35-9b-equational-theory-sair-syncards-e-notes50m-card3-ntp-kl-0p01"processor = AutoProcessor.from_pretrained(model_id, token=True)model = AutoModelForImageTextToText.from_pretrained( model_id, token=True, dtype="auto", device_map="auto")The Stage II prompts, the sampling settings (temperature 1.0, top-p 0.95, top-k 20, presence penalty 1.5, 24,576 output tokens)
and the grading pipeline are documented in the Space and in the repository's recipe/equational_internalization/evaluation/GRADING.md.