GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zhiqilicg
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
zhiqilicg
Model Tree
Qwen/Qwen3.6-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
LoRA SFT from the original model, using 313 verified records: 182 software-repair, 98 terminal and 33 chat. Task records were generated by Qwen3.6-27B and the unmodified Qwen3.6-35B-A3B. 19 steps; constant learning rate 1e-4.
Base revision: 995ad96eacd98c81ed38be0c5b274b04031597b0. Run: hsftc_e1p1F. Checkpoint step within the run: 19.
LoRA r=8, alpha=32, dropout=0.05; all-linear language-model modules, including routed experts. BF16 weights, no weight quantization. Global batch 16; context 131072 tokens; random seed 42. Non-thinking mode. Tokenizer and chat template are unchanged from the pinned base; the training template is qwen3_5.
Software-repair tasks: nebius/SWE-rebench-V2 (CC-BY-4.0). Terminal tasks: camel-ai/seta-env and open-thoughts/TaskTrove v5.1 (Apache-2.0). Chat: CC-BY-4.0 records from nvidia/Llama-Nemotron-Post-Training-Dataset, SFT/chat. Generated records are filtered by execution verification, quality checks, deduplication and benchmark overlap checks. Revisions, selection code and record inventories are in the GitHub repository.
| Benchmark | Score (%) |
|---|---|
| SWE | 25.49 |
| HLE | 13.12 |
| ASI | Pending |
| TB2 | 40.0 ± 4.54 |
SWE uses HARD-51; HLE uses the full 2500-question set; ASI uses the local 57-task setup; TB2 is the five-run mean ± sample standard deviation. Pending evaluations remain pending in the report. This model is provided for agent-training research and reproduction of the reported experiments; detailed settings and observed failure cases are in GitHub.
Each released adapter is complete relative to the original base. Merge it directly into that base; do not stack the SFT adapter first. The merger handles per-expert LoRA matrices and fused MoE base tensors. Load the merged output with Transformers or vLLM.
bash
hf download Qwen/Qwen3.6-35B-A3B --revision 995ad96eacd98c81ed38be0c5b274b04031597b0 --local-dir ./basehf download zhiqilicg/qwen36-35b-a3b-hybrid-sft-lora --local-dir ./adapterpython ./adapter/merge_lora.py --base ./base --adapter ./adapter --out ./mergedThe merge uses CPU, torch and safetensors, and writes BF16 model weights. Adapter configuration, ms-swift training arguments and the merge script are included. License: Apache-2.0.
Weight SHA256: 8c61f5faf6169b1734e14cc301bff26ef853664717972f20df6105d46b94d308.