GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
skyline02
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
skyline02
Model Tree
Qwen/Qwen3.6-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
Base: Qwen/Qwen3.6-35B-A3B, revision 995ad96eacd98c81ed38be0c5b274b04031597b0.
This immutable adapter is selected from completed training on the independent
development set. The earlier revisions in this repository remain pilot artifacts.
Actual training examples: 31; optimizer updates: 8; selected development completion loss: 0.28617432713508606. Training time: 167.6 seconds; peak reserved GPU memory: 76.61 GiB. LoRA rank 16, alpha 32, dropout 0.05. Base and inference use BF16; saved adapters use FP32. Routed experts, routers and vision components are frozen. Training context is 8192; seed 42; microbatch 1, accumulation 4; completion-only supervision; learning rate 5e-5 for SFT and 1e-5 for execution-filtered continued SFT. Detailed run metadata is attached.
The expanded corpus has 396 verified SFT records, 119 independent development records and 69 independent execution-feedback tasks. Data derives from permissively licensed public repair commits and tests. Stable failing-before/passing-after checks, regression checks, historical license notices, patch deduplication and frozen benchmark-input overlap checks precede training. The second stage generates new patches from the selected SFT adapter, executes them, and continues training with 30% original-data replay by supervised tokens. Benchmark answers, reference patches, hidden tests and feedback are not training or checkpoint-selection inputs.
The corpus remains Python-focused and commit-derived source excerpts disclose repair locations. Input overlap checks do not prove absence of semantic or base pretraining contamination. Development loss is not a benchmark score. No benchmark improvement is claimed before the complete frozen three-version comparison.
python
import torchfrom transformers import AutoModelForImageTextToText, AutoTokenizerfrom peft import PeftModelbase_id = "Qwen/Qwen3.6-35B-A3B"revision = "995ad96eacd98c81ed38be0c5b274b04031597b0"base = AutoModelForImageTextToText.from_pretrained( base_id, revision=revision, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, "skyline02/qwen36-35b-a3b-se-lora-rest", revision="ADAPTER_REVISION")model = model.to(dtype=torch.bfloat16).eval()tokenizer = AutoTokenizer.from_pretrained(base_id, revision=revision)Replace ADAPTER_REVISION with the immutable commit recorded in the experiment.
Keep the official tokenizer/chat template and enable_thinking/preserve_thinking
settings fixed within each benchmark. The full approximately 35B parameter weights
must fit in memory; the approximately 3B active parameters do not determine weight
storage. Validated training: torch 2.12.1, transformers 5.18.0, PEFT 0.21.2,
Python 3.12. Evaluation serving: vLLM 0.30.0 and BF16. The model is public and ungated.