GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
violetxi
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
violetxi
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Full SFT NLL plus 0.1 × KL(base || student) over the full
vocabulary at sampled assistant prediction positions in the frozen replay pool.
The KL term uses the mean of trajectory means, with up to
128 sampled positions per trajectory. Its replay data
is additional to the supervised note/recall token budget above.
Final fixed-replay diagnostic KL: 0.01216539;
this is not a benchmark score. Details and provenance hashes are in
training_summary.json.
Use a recent Qwen3.5-compatible Transformers/vLLM installation. The frozen evaluation pipeline uses vLLM 0.21.0; publication verification records the actual Transformers version in its verification receipt.
python
from transformers import AutoModelForImageTextToText, AutoTokenizermodel_id = "violetxi/qwen35-9b-wmrl-v4-scale70-30m-kl-0p1"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="bfloat16", device_map="auto", use_safetensors=True)
bash
vllm serve violetxi/qwen35-9b-wmrl-v4-scale70-30m-kl-0p1 --dtype bfloat16 --max-model-len 65536
Agentic 5-turn: 9.79% mean rubric score.
Agentic 20-turn: 12.64% mean rubric score.
Closed-book results are in evaluation_summary.json.
The agentic protocol uses 250 tasks, four samples/task, thinking, the frozen DMS tools/documents and GLM judge. Closed-book evaluation uses the original probe sampling and scoring. This release represents one training seed.
Base revision: c202236235762e1c871ad0ccb60c8ee5ba337b9a. Model parameter/index validation,
file sizes and SHA-256 hashes are in publication_manifest.json.