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
The fixed dataset contains 99,998,917 supervised tokens per epoch before causal shifting: 69,999,985 notes tokens (157,434 notes) and 29,998,932 assistant trajectory tokens (10,306 note-conditioned trajectories). All note tokens are supervised; trajectory loss covers assistant tokens. After shifting, each epoch has 99,993,091 supervised tokens; two epochs expose 199,986,182. KL replay is additional to this supervised budget.
Objective: notes CPT + assistant-masked note-conditioned trajectory SFT + 0.1 × KL(base || student). The reference is the frozen pinned base. KL uses the full vocabulary and the mean of session means, sampling up to 128 assistant prediction positions per original-session replay. The fixed replay pool has 2,048 sessions, context limit 8,192, seed 731, and 24,544 draws across training.
Training starts from the pinned base: c202236235762e1c871ad0ccb60c8ee5ba337b9a. Context rows are 16,384 tokens; global batch 8; gradient accumulation 1; two B200 GPUs with four rows each; 3,068 updates; learning rate 5e-6, cosine schedule, warmup ratio 0.03, seed 0. Each epoch contains 12,268 real rows plus four zero-loss padding rows. Full training and replay provenance is in training_summary.json.
Final held-out KL(base || student) = 0.008930197923, over 256 fixed held-out sessions / 32,768 prediction positions. It measures drift from the base, not downstream task accuracy. Initial and epoch-end diagnostics are included in evaluation_summary.json.
Agentic and closed-book benchmark evaluation has not been run for this checkpoint. No downstream benchmark scores are claimed. No fresh GPU inference was performed after this serving export.
This repository contains the complete composite model in four safetensors shards, matching config, tokenizer/chat template, image/video processor metadata, and generation settings. The trained text checkpoint was exported from FP32 to the pinned base's BF16 serving dtype. All 427 trained tensors were checked for exact equality to that conversion; all retained vision/auxiliary tensors were checked against the pinned base. Every exported tensor is finite. Vision and auxiliary components were not trained or evaluated for multimodal quality. No adapter or local reconstruction is required.
python
from transformers import AutoModelForImageTextToText, AutoTokenizermodel_id = "violetxi/qwen35-9b-harvey-v4-notes-conditioned-100m-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-harvey-v4-notes-conditioned-100m-kl-0p1 --dtype bfloat16 --max-model-len 65536
Use a Qwen3.5-compatible Transformers/vLLM installation. File sizes and SHA-256 hashes are recorded in publication_manifest.json. This release is one training seed.