GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
omrisap
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
omrisap
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Full input vocabulary is preserved. Output logits are projected ONLY onto
the 12 allowed grid tokens (digits, newline, end-of-answer). Unrestricted native
generate() is NOT equivalent. Use the included grid_solver.py helper:
python
import json, torchfrom transformers import AutoProcessor, Qwen3_5ForConditionalGenerationfrom huggingface_hub import hf_hub_downloadfrom grid_solver import GridSolver, generate_gridrepo = "omrisap/qwen35-4b-grid-a-50k"processor = AutoProcessor.from_pretrained(repo)contract = json.load(open(hf_hub_download(repo, "solver_contract.json")))base = Qwen3_5ForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16).cuda()solver = GridSolver(base, contract["identity"]["allowed_ids"]).eval()# inputs: processor-generated tensors for hybrid alternating demo messages,# query input, enable_thinking=False and add_generation_prompt=True.# generated_ids, terminated = generate_grid(solver, inputs, max_new_tokens=931)The model root contains standard safetensors, config, tokenizer and processor.
training/checkpoint_latest.pt additionally preserves optimizer/RNG state for
exact resumption using the original Phase 8 code and registered dataset.
Only load this PyTorch checkpoint if you trust its source.
import json, torchfrom transformers import AutoProcessor, Qwen3_5ForConditionalGenerationfrom huggingface_hub import hf_hub_downloadfrom grid_solver import GridSolver, generate_gridrepo = "omrisap/qwen35-4b-grid-a-50k"processor = AutoProcessor.from_pretrained(repo)contract = json.load(open(hf_hub_download(repo, "solver_contract.json")))base = Qwen3_5ForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16).cuda()solver = GridSolver(base, contract["identity"]["allowed_ids"]).eval()# inputs: processor-generated tensors for hybrid alternating demo messages,# query input, enable_thinking=False and add_generation_prompt=True.# generated_ids, terminated = generate_grid(solver, inputs, max_new_tokens=931)