Langboat open-sources the Mengzi3 model family

A Chinese-focused family

On 29 March 2024 Langboat released the Mengzi3 model family in two sizes, 8B and 13B ^[1][2]. The first release covered the base weights; the company said the 13B chat version was still to come ^[1]. Both base models are published under the Apache-2.0 licence and are free for commercial use ^[1].

Where the weights went

The models were put on Hugging Face, ModelScope, Wisemodel and the MindSpore MindFormers collection at the same time ^[1]. Publishing to several hubs at once lets a user pick the platform their existing tooling already talks to, and the MindFormers listing places the weights inside the MindSpore ecosystem rather than only on the hubs most commonly used with other frameworks.

What the benchmarks said

Langboat reported that Mengzi3-13B-Base beat Baichuan2-13B, Qwen-14B and Skywork-13B on MMLU and CMMLU in 5-shot tests ^[1]. MMLU is a multiple-choice benchmark that samples questions from many academic subjects and is predominantly English; CMMLU is its Chinese counterpart, built from questions set in a Chinese context. "5-shot" means the prompt contains five worked examples before the question being scored, with no extra training — a standard way of comparing base models, which are trained only to continue text. The figures were reported by the company itself, and the three models it was compared against are open-weight models from Chinese teams, so the comparison is reasonably like for like on language coverage. The notable part is the size gap: a 13B model was reported ahead of a 14B model on both benchmarks ^[1].

Positioning

Langboat positioned Mengzi3 around Chinese ability while keeping multilingual competence ^[1]. That is a narrower claim than "best overall model": the target is a model that handles Chinese well without losing the ability to work in other languages.

The chat versions came later

The 8B chat model reached Hugging Face on 14 September 2024, nearly six months after the base weights ^[2]. Base and chat weights are different products. A base model is trained to predict the next token and serves mainly as a starting point for further training and for benchmark comparisons; a chat model needs additional post-training and evaluation before it can be released, and that work is where much of the elapsed time goes.

Why it matters

Three things make the release worth recording: a permissive licence that allows commercial use without a separate agreement ^[1]; distribution across four hubs at once ^[1]; and a 13B model reported ahead of a 14B competitor on MMLU and CMMLU ^[1]. Read together, they describe a release pattern that has become common for open-weight models: ship the base weights and the licence first, then follow with chat versions once post-training is finished ^[2].