29B total / 4B active (MoE)
262,144 tokens native; extensible to 524,288 tokens
Apache-2.0
Apache-2.0 task-specialized fine-tune of Xing4.0-29B-A4B for schema-guided intent recognition and structured information extraction. Xing4.0-29B-Intent uses a Unified Schema-driven Semantic Parsing protocol that represents intent recognition and slot filling, named-entity recognition, relation extraction, and event extraction through the same category-plus-args JSON structure. Users provide candidate categories, descriptions, and extractable field definitions; the model jointly selects one or more categories and fills one or more structured argument instances. The inherited Xing4.0 architecture is a 40-layer, 29B-total/4B-active mixture of experts combining manifold-constrained Hyper-Connections, Multi-head Latent Attention, and Multi-Token Prediction. Its configuration contains a 3,584-dimensional hidden state, two initial dense layers, 64 routed experts plus one shared expert, four routed experts selected per token, and a 262,144-token native context window advertised as extensible to 524,288 tokens. The release focuses on Chinese and English inference use cases and includes a Python XingIntent SDK for schema definitions, prompt construction, multi-turn conversations, synchronous and asynchronous requests, constrained decoding, and structured response parsing. Its companion Xing4.0-29B-Intent-oss dataset contains 2,753,000 records spanning 61 reconstructed public datasets, China Telecom smart-device-assistant annotations, 19 domains, 52 languages, and four task families: 823,000 named-entity examples, 905,000 relation examples, 56,000 event examples, and 969,000 intent/slot examples. On a 6,000-sample test set drawn from 48 public datasets, China Telecom reports an overall score of 58, with 69 for category selection, 40 for event extraction, 72 for named-entity recognition, 38 for relation extraction, and 81 for intent understanding. These are macro-aggregated field-value Micro-F1 scores on a 0–100 scale, not business accuracy percentages. The publisher reports a higher overall score than Gemini-3.6-Flash at 49, Qwen3.8-Max at 46, Kimi-K3 at 45, and GPT-5.6-Sol at 45 under its evaluation. The BF16 release occupies approximately 62.4 GB across 41 safetensors shards; official FP8 and IQ4_NL GGUF variants are released alongside it. Deployment follows the custom-code Xing4.0 base through vLLM or SGLang and exposes an OpenAI-compatible API.