XingChen: Xing4.0 29B A4B (free)

XingChen: Xing4.0 29B A4B (free) has a clean measured fingerprint that currently sits alone: no other measured model shares its signature.

Xing4.0 comes from China Telecom's AI company, which publishes as XingChen-AGI and renamed its TeleChat models Xing. Its released tokenizer, run over the same 50 strings offline, returns the served counts on all 50: a SentencePiece vocabulary of 131,072 pieces, TeleChat3's, with 131,070 of them at the same positions, and a signature new to the dataset. The architecture is DeepSeek's published design. Its config.json uses 41 of DeepSeek-V3's setting names among its 47, with the sizes scaled down, among them MLA attention, a sigmoid expert router and one multi-token prediction layer, and three more are DeepSeek-V4's mHC settings at DeepSeek's values; its configuration code opens with DeepSeek-V3's own lines. Asked who made it in English, the model names Google Brain, or OpenAI with thinking on; asked in Chinese, it names China Telecom.

Xing4.0-29B-A4B repository (config.json: vocab_size 131072, 64 routed experts, hc_mult 4)Xing4.0-29B-A4B configuration code (opens with DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP)TeleChat3-36B-Thinking repository (tokenizer.model, 131,072 pieces)DeepSeek-V3 repository (config.json)DeepSeek-V4-Flash repository (config.json: hc_mult 4, hc_sinkhorn_iters 20, hc_eps 1e-6)