--- library_name: transformers license: apache-2.0 base_model: Qwen/Qwen3.5-35B-A3B base_model_relation: finetune pipeline_tag: text-generation language: - en - zh tags: - apus-reflex - decision-model - structured-output - mixture-of-experts - bf16 --- # APUS Reflex 35B-A3B > **Branding note:** This model family is now **APUS Reflex**. Some weight filenames, script/module names, served-model aliases, and the separate evaluation dataset ID retain their historical identifiers for download and runtime compatibility; these do not indicate different model versions. [English](README.md) | [中文](README.zh-CN.md) · [Collection](https://huggingface.co/collections/apus-ailab/apus-reflex-6ab1ee888eb002fcdd3a2825) · [Model family](https://huggingface.co/apus-ailab/APUS-Reflex) · [Technical Report](https://huggingface.co/apus-ailab/APUS-Reflex/blob/main/TECHNICAL_REPORT.md) · [Runtime](RUNTIME.md) A Qwen3.5 MoE decision model for choosing browser actions, selecting workflow steps, and judging natural-language criteria. Give the model a shared state and a set of candidate actions; the included decision runtime returns a distribution over those candidates. This repository contains **35B-A3B checkpoint-5949 merged BF16 weights**. It is a standalone model with root-level Hugging Face configuration and weights, requiring no separate LoRA adapter. The native merged model scores **71/80 (88.75%)** at its full 40-layer depth on the Frozen80 development panel. ## Choose a decision budget | 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](RUNTIME.md) for the portable interface and its exact input contract. ## Download and run ```bash python -m pip install huggingface_hub hf download apus-ailab/APUS-Reflex-35B-A3B --local-dir ./APUS-Reflex-35B-A3B cd APUS-Reflex-35B-A3B python -m pip install -r requirements.txt python examples.py . --device cuda:0 --effort high ``` The examples use the portable reference runtime; its current acceptance status is recorded in [RUNTIME.md](RUNTIME.md). For reproducibility, pass an immutable repository commit with `hf download --revision `. 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.model tokenizer = runtime.tokenizer ``` Use 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. ## Evaluation and training | 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](https://huggingface.co/datasets/apus-ailab/APUS-OpenJev-Eval-Frozen80) 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](training.md). Numerical merge differences, experimental-mode boundaries and validation status are collected in [RUNTIME.md](RUNTIME.md). ## GGUF and MLX versions Quantized versions of this model (Frozen80 71/80) for Ollama / llama.cpp / LM Studio and Apple Silicon Macs: [GGUF collection](https://huggingface.co/collections/apus-ailab/apus-reflex-gguf-6ab39d5e724c4d8a1021198f) · [MLX collection](https://huggingface.co/collections/apus-ailab/apus-reflex-mlx-6ab39d5fc988a1b1cb89dfc8). | Version | Frozen80 | Decisions = BF16 | |---|---:|---:| | [GGUF Q8_0](https://huggingface.co/apus-ailab/APUS-Reflex-35B-A3B-GGUF) | 71/80 | 80/80 | | [GGUF Q4_K_M](https://huggingface.co/apus-ailab/APUS-Reflex-35B-A3B-GGUF) | 72/80 | 79/80 | | [MLX 8bit](https://huggingface.co/apus-ailab/APUS-Reflex-35B-A3B-MLX-8bit) | 71/80 | 80/80 | | [MLX 4bit](https://huggingface.co/apus-ailab/APUS-Reflex-35B-A3B-MLX-4bit) | 70/80 | 78/80 | ```bash ollama run hf.co/apus-ailab/APUS-Reflex-35B-A3B-GGUF:Q8_0 --think=false ``` Ollama needs thinking disabled (`--think=false`, or `"think": false` in the API). Per-question results and usage are in each repository. ## Series and license The [Collection](https://huggingface.co/collections/apus-ailab/apus-reflex-6ab1ee888eb002fcdd3a2825) groups the independent 4B, 9B and 35B-A3B repositories. Each model has its own checkpoint identity, evaluation and download counter. The [family report](https://huggingface.co/apus-ailab/APUS-Reflex/blob/main/TECHNICAL_REPORT.md) provides the shared research context. We thank the [Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) team. Model licensing follows Apache-2.0; retain the accompanying license and provenance files when redistributing. **Authors:** gumpcheng ([xDAN2099](https://huggingface.co/xDAN2099)), zhangxu, [APUS AI-LAB](https://github.com/APUS-AI-Lab).