GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
apus-ailab
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
apus-ailab
Model Tree
Qwen/Qwen3.5-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
| Mode | Decoder layers | Intended use |
|---|---|---|
effort="high" | 40 | Primary release mode for decisions and text generation |
effort="low" | 20 | Experimental shallow decision exit, selected explicitly |
The decision interface accepts 2–16 request-specific candidates. Applications can use the returned candidate IDs to dispatch actions or build structured workflow results. See runtime instructions and validation status for the portable interface and its exact input contract.
bash
python -m pip install huggingface_hubhf download apus-ailab/APUS-OpenJev-v1-35B-A3B --local-dir ./APUS-OpenJev-v1-35B-A3Bcd APUS-OpenJev-v1-35B-A3Bpython -m pip install -r requirements.txtpython examples.py . --device cuda:0 --effort highThe examples use the portable reference runtime; its current acceptance status is recorded in RUNTIME.md. For reproducibility, pass an immutable repository commit with hf download --revision <commit>.
To access the underlying full-depth Transformers model using the validated loader:
python
from openjet_runtime import OpenJet runtime = OpenJet.from_pretrained(".", device="cuda:0", dtype="bfloat16")model = runtime.modeltokenizer = runtime.tokenizerUse the pinned dependencies in requirements.txt. The bundled loader preserves sensitive checkpoint parameters in FP32 while loading the main weights in BF16. It also provides the 20-layer exit and candidate-distribution interface. A rough planning estimate is 75 GB or more GPU memory for BF16 inference, varying with prompt length, batching and runtime. This release's native evaluation used an RTX PRO 6000 with 96 GB; the estimate is not a tested minimum.
| This checkpoint's merged model | Correct / 80 | Accuracy |
|---|---|---|
| Native full 40 layers | 71 / 80 | 88.75% |
| Native experimental 20 layers | 63 / 80 | 78.75% |
The Frozen80 panel covers Browser, HelpSteer3, BoolQ, MNLI and attribute decisions. These results come from the merged checkpoint-5949 native GPU evaluation, not the source adapter's larger evaluation panels. Frozen80 is a reused engineering development panel, not a blind benchmark or an end-to-end browser success rate.
Training used 5,949 curriculum records from 3,898 parent groups, with LoRA on attention and shared experts while routed experts and the router remained frozen. See training details. Numerical merge differences, experimental-mode boundaries and validation status are collected in RUNTIME.md.
Quantized versions of this model (Frozen80 71/80) for Ollama / llama.cpp / LM Studio and Apple Silicon Macs: GGUF collection · MLX collection.
| Version | Frozen80 | Decisions = BF16 |
|---|---|---|
| GGUF Q8_0 | 71/80 | 80/80 |
| GGUF Q4_K_M | 72/80 | 79/80 |
| MLX 8bit | 71/80 | 80/80 |
| MLX 4bit |
bash
ollama run hf.co/apus-ailab/APUS-OpenJev-v1-35B-A3B-GGUF:Q8_0 --think=falseOllama needs thinking disabled (--think=false, or "think": false in the API). Per-question results and usage are in each repository.
The Collection groups the independent 4B, 9B and 35B-A3B repositories. Each model has its own checkpoint identity, evaluation and download counter. The family report provides the shared research context.
We thank the Qwen/Qwen3.5-35B-A3B team. Model licensing follows Apache-2.0; retain the accompanying license and provenance files when redistributing.
Authors: gumpcheng (xDAN2099), zhangxu, APUS AI-LAB.
| 70/80 |
| 78/80 |