GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
XAILab-CyberSpark
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
XAILab-CyberSpark
Model Tree
Qwen/Qwen3.5-0.8B
Input Modalities
Output Modalities
Supported Functionality
本模型以 Qwen3.5-0.8B 为基座，在一次前向计算中处理车辆状态和多组候选问题／选项。问题共享 State 但彼此隔离；pointer head 给每组调用方提供的候选项打分，不进行自回归 JSON 生成。LoRA rank 16、alpha 32、dropout 0.05；pointer head 维度 256。
bash
pip install -r requirements.txtpython inference.py --record example_record.json --device cuda --max-state-tokens 768# 本地基座权重：附加 --base-model /path/to/Qwen3.5-0.8B公开接口为 state 加若干 question／options。softmax 仅在每组给定选项内比较，不是校准过的服务推荐概率。推理需要仓库的自定义 src/ 代码和另行取得的 Qwen 基座权重。原型还使用调用方侧的候选生成器、车辆服务 schema 与确定性 DAG 组装；相关服务 schema/API 资产不在此仓库中，因此公开示例不能复现完整 DAG 基准。
训练采用覆盖 12 类车辆控制意图的中文合成指令。在项目完整内部流水线上，IID/OOD 留出合成集各 200 条，exact DAG match 分别为 88.0% / 76.5%；这些不是公开示例单独运行的得分。OOD 误激活 1/46（2.17%），高于项目的 2% 目标。其他域无正例训练；候选召回和确定性组装也会限制端到端结果。训练／评估数据和车辆 API 目录未发布。合成测试不能证明真实 ASR、未见意图或车辆控制可靠性；实际推荐需要独立评估和概率校准。
授权与署名： 模型权重及原创推理代码 © 2026 OpenSparX（上海塞伯火种人工智能有限公司），Apache-2.0。基座模型的署名与条款见 Qwen/Qwen3.5-0.8B。仓库不含私有凭据、真实车辆命令分发器或 QNN 编译产物。
Developer: OpenSparX (上海塞伯火种人工智能有限公司). License: Apache-2.0. This release contains the floating-point LoRA adapter, pointer decision head, and generic State/Question inference code. Download the Qwen3.5-0.8B base weights separately. The model scores isolated option sets in one shared-State forward pass; it is not an autoregressive JSON generator.
bash
pip install -r requirements.txtpython inference.py --record example_record.json --device cuda --max-state-tokens 768# Offline base-model copy: add --base-model /path/to/Qwen3.5-0.8BThe example shows the public model-level interface: state plus a list of question/options records. Scores are softmax values over each caller-provided option set, not calibrated service probabilities. The original cabin prototype also used a caller-side candidate generator, vehicle-specific service schema, and deterministic DAG assembly. Those service-schema and API assets are not part of this release, so this generic example is not a reproduction of the full DAG benchmark.
The model uses a LoRA rank-16 adapter, alpha 32, dropout 0.05, and a 256-dimensional pointer head. It was trained on synthetic Chinese cabin commands covering 12 vehicle-control intents. On the project’s full internal pipeline, held-out synthetic IID/OOD sets of 200 records each reached 88.0% / 76.5% exact DAG match. These are pipeline metrics, not scores from this standalone example. The OOD false-activation rate was 1/46 (2.17%), above the project’s 2% target. Other domains did not have positive training examples; candidate recall and deterministic assembly constrain end-to-end performance. The training/evaluation data and proprietary vehicle API catalog are not included.
src/ code and a separate Qwen base model. Do not call generate() to obtain JEV decisions.Model weights and original JEV inference code: © 2026 OpenSparX, Apache-2.0. Base-model attribution and terms: Qwen/Qwen3.5-0.8B. This repo contains no real vehicle command dispatcher or private credentials.
pip install -r requirements.txtpython inference.py --record example_record.json --device cuda --max-state-tokens 768# 本地基座权重：附加 --base-model /path/to/Qwen3.5-0.8Bpip install -r requirements.txtpython inference.py --record example_record.json --device cuda --max-state-tokens 768# Offline base-model copy: add --base-model /path/to/Qwen3.5-0.8B