GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zwright
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
zwright
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
On identical frozen browser tasks, the unmodified base scored 0/4 development
and 0/9 sealed-final overall success. The adapter scored 4/4 and 9/9,
respectively. Functional tool-call and exact final-answer rates moved by the
same amount. Literal tool-call argument exactness remained 0/4 and 0/9 because
the adapter emits the evaluator's narrowly permitted trailing headline
search refinement.
The adapter evaluation was the governed one-shot sealed action. The matched base control was run post-hoc after weights and recipe were frozen and is clearly labeled as such in the evidence. Full reports, paired per-task scores, training curve, manifests, checksums, and limitations are in the associated GitHub evidence capsule.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelrevision = "cdbee75f17c01a7cc42f958dc650907174af0554"base_id = "Qwen/Qwen3-4B-Instruct-2507"adapter_id = "zwright/qwen3-4b-hermes-flight-recorder-browser-lora"tokenizer = AutoTokenizer.from_pretrained(base_id, revision=revision)base = AutoModelForCausalLM.from_pretrained(base_id, revision=revision)model = PeftModel.from_pretrained(base, adapter_id)
Use the base model's native chat template and tool definitions. A deterministic external router should activate this adapter only for the evaluated browser scope; model text must not choose its own adapter, tools, or write authority.
1e-4 with batch size 1 and gradient accumulation 4.0.0459748519.aeeb93af9a0b35a7c685942c320ae2465fd06edc73b9ee85d416918208a2e5e6.Do not merge or stack this adapter with another adapter without treating the composition as a new candidate with its own immutable hash and evaluation.
cdbee75f17c01a7cc42f958dc650907174af0554.64f85b198065d0fba2a81f37e10ed68161ce2c19a754c7100e67e0ca2ee9c326.a1495c41471e0b09230fcf1d9ef2556e88ca856dcab489aa110b302b669d3dc7.The base model and this adapter are distributed under Apache-2.0. Users remain responsible for reviewing upstream terms and validating behavior for their own deployment.