Nanbeige introduces a model for coding and office tasks

Nanbeige targets coding and office tasks with this release. The model card separates two reasoning controls. One switches reasoning on for the current response; it is enabled by default. The other retains reasoning from earlier assistant turns. Nanbeige recommends retaining it for tool use across multiple turns, coding and office workflows, but not for ordinary chat. Supplying tools selects the tool-use template.

The separate Base repository targets developers who want further training for general or specialized agents. Its Python example shows text completion: after loading the tokenizer, which converts text into model inputs, and the model, it turns an unfinished sentence into model inputs and generates a continuation. It removes the original input tokens before decoding and printing the new text. The repository also provides SGLang and vLLM serving commands with four-way tensor parallelism. These are supplied examples, not a model run performed for this article.

The configuration file gives developers concrete implementation details: it names LlamaForCausalLM as the architecture, specifies 52 layers and records bfloat16 as the numerical format. It also records Transformers version 4.51.1. These settings describe the supplied configuration; the file contains no measured response time or measured GPU-memory requirement.

MarkTechPost’s report covers the release and links both repositories, giving readers direct paths to the developer’s documentation.