Xingchen 4.0 29B
A domestically developed open-source large language model using a Mixture-of-Experts (MoE) architecture, with 29 billion total parameters and 4 billion activated per inference. The model places particular emphasis on Agent and coding capabilities, supports private deployment, and is compatible with mainstream Agent frameworks and domestic AI chips.
Timeline (last 90 days)
Xingchen 4.0 29B is a domestically developed open-source model that adopts a MoE architecture with 29B total parameters, activating only 4B parameters at a time.
Xingchen 4.0 29B focuses on enhancing Agent and Coding capabilities at the 29B scale and supports privatized deployment.
Xingchen 4.0 29B is compatible with mainstream Agent and chat frameworks such as Claude Code, Cline, and Hermes.
After 4-bit quantization, Xingchen 4.0 29B occupies approximately 15GB of VRAM on a single GPU and can be deployed on consumer gaming graphics cards.
Xingchen 4.0 29B is compatible with fine-tuning tools such as MindFormers and LLaMA-Factory, as well as inference frameworks such as SGLang and vLLM.
Xingchen 4.0 29B has completed end-to-end adaptation for multiple domestically produced chips including Ascend based on Zhiyuan FlagOS, enabling cross-chip architecture migration and one-click deployment.
The computational load of each inference pass of Xingchen 4.0 29B is roughly equivalent to that of a 4B dense model, as VRAM usage and compute requirements are determined by the number of activated parameters rather than the total parameter count.
Xingchen 4.0 is built on a data foundation of hundreds of trillions of tokens, adopts architectural designs including MLA, MTP, and MHC, and undergoes staged long-context training from 4K to 32K, 128K, and 256K.
Xingchen 4.0 conducts multi-expert reinforcement learning training separately in the Agent, Coding, and Reasoning directions, and integrates the capabilities of different experts through MoPD.
Xingchen 4.0 29B has been released on open-source communities including GitHub, Hugging Face, and ModelScope.
1 more timeline events
All Facts (11)
In a Coding practical test, when fixing a date calculation bug in a Git repository, Xingchen 4.0 followed engineering best practices by first gathering evidence to diagnose the problem, controlling the scope of modifications, and verifying with tests.
50%UnverifiedXingchen 4.0 29B focuses on enhancing Agent and Coding capabilities at the 29B scale and supports privatized deployment.
50%UnverifiedXingchen 4.0 29B is compatible with mainstream Agent and chat frameworks such as Claude Code, Cline, and Hermes.
50%UnverifiedAfter 4-bit quantization, Xingchen 4.0 29B occupies approximately 15GB of VRAM on a single GPU and can be deployed on consumer gaming graphics cards.
50%UnverifiedXingchen 4.0 29B is compatible with fine-tuning tools such as MindFormers and LLaMA-Factory, as well as inference frameworks such as SGLang and vLLM.
50%UnverifiedXingchen 4.0 29B has completed end-to-end adaptation for multiple domestically produced chips including Ascend based on Zhiyuan FlagOS, enabling cross-chip architecture migration and one-click deployment.
50%UnverifiedThe computational load of each inference pass of Xingchen 4.0 29B is roughly equivalent to that of a 4B dense model, as VRAM usage and compute requirements are determined by the number of activated parameters rather than the total parameter count.
50%UnverifiedXingchen 4.0 is built on a data foundation of hundreds of trillions of tokens, adopts architectural designs including MLA, MTP, and MHC, and undergoes staged long-context training from 4K to 32K, 128K, and 256K.
50%UnverifiedXingchen 4.0 conducts multi-expert reinforcement learning training separately in the Agent, Coding, and Reasoning directions, and integrates the capabilities of different experts through MoPD.
50%UnverifiedXingchen 4.0 29B is a domestically developed open-source model that adopts a MoE architecture with 29B total parameters, activating only 4B parameters at a time.
50%UnverifiedXingchen 4.0 29B has been released on open-source communities including GitHub, Hugging Face, and ModelScope.
50%