GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
RichWoollcott
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
RichWoollcott
Model Tree
Base
Qwen/Qwen3-4B-Instruct-2507
Adapter
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
The LoRA adapter (r=16, α=32, attention+MLP targets) behind DCL-Qwen3-4B-Instruct-2507, a fine-tune of Qwen/Qwen3-4B-Instruct-2507 that authors and repairs DCL (Russell East's Declarative Capability Language). See the main model card for the evaluation table, the required serving prompt, and limitations.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelimport torchm = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507",dtype=torch.bfloat16, device_map="auto")m = PeftModel.from_pretrained(m, "RichWoollcott/DCL-Qwen3-4B-Instruct-2507-LoRA")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
Note: the adapter was trained against the 4-bit quantised base (QLoRA); applied to the full-precision base as above, behaviour matches the released merged model to within quantisation noise — the frozen-exam receipts were graded on the merged Q4_K_M build.
Trained on a private corpus of 507 compiler-verified, fully synthetic rows with Unsloth + TRL (PEFT 0.19.1) — kept private to protect the integrity of the frozen evaluation exam. Apache-2.0; base LICENSE retained.