通义千问开放 Qwen 权重

2023 年 8 月 3 日，阿里云团队把 Qwen-7B 与 Qwen-7B-Chat 放上 ModelScope 和 Hugging Face。对开发者而言，关键动作不是又多了一个聊天网页，而是模型从网页背后走到了本地目录：代码与权重可以下载，基础模型可以继续训练，Chat 版本可以在私有环境里测试。中文大模型不再只能通过厂商 API 被远程调用，它开始拥有一个能够被社区共同指向的公开检查点。

首版 Qwen-7B 报告使用超过 2.2 万亿 token 的预训练语料，覆盖中文、英文、代码和其他多语言文本。基础模型学习续写与表示，Chat 版本再经过指令与对齐训练。这里的中文能力不只是“数据里含有中文”这么简单：分词器如何切分汉字和词组，会直接影响固定上下文里能装进多少内容，也影响训练与推理的 token 成本。对于此前常要忍受英文模型翻译腔、中文知识薄弱或分词低效的团队，这些底层选择会在每一次提示中出现。

分发方式同样重要。ModelScope 为国内网络与阿里生态提供入口，Hugging Face 则把权重放进全球开发者已经使用的模型工具链。仓库给出 Transformers、量化和微调示例，之后又陆续补充 Int4、Int8 与 LoRA 等路径。首日发布并没有一次性完成所有工具，但它建立了一个循环：权重被下载，问题出现在 issue 里，适配代码被更新，下一版模型再被同一批用户接住。

许可为这个循环画了一条明确但不算宽松到底的边界。研究使用被允许；当时条款允许月活用户低于一亿的商业主体使用，达到或超过门槛则需另行申请。它应被称为开放权重和有条件商业许可，而不是 Apache 式无限制开源。对多数个人、高校和初创公司，这道门槛并不妨碍试验；对超大平台，许可证本身就是采用决策的一部分。

发布方基准覆盖中英文理解、代码和数学等任务，可以说明团队想证明什么，却不能替代独立评测。更可观察的变化发生在使用方式上：企业可以把中文基座放进内网，研究者可以在相同权重上复现实验，开发者可以围绕一个公开版本报告缺陷。模型是否在某张表上第一会迅速过时，而这种可持续分发关系一旦形成，后续迭代就不再是彼此孤立的新闻。

后来 Qwen 扩展到更多参数规模、视觉、音频、代码与数学模型，许可也继续调整。那些进展不应倒灌到 8 月 3 日。首日真正发生的事情更具体：一组以中文和英文为重点训练的 7B 权重，同时抵达两个模型社区；从那天起，“通义的下一版模型”不再只是一项云产品更新，也成为开发者会主动等待、下载并比较的公共事件。

On 3 August 2023, Alibaba Cloud’s team placed Qwen-7B and Qwen-7B-Chat on ModelScope and Hugging Face. For developers, the important event was not another chat page. The model moved from behind a service into a local directory: code and weights could be downloaded, the base model could be further trained, and the chat model could be tested inside a private environment. A Chinese large language model was no longer available only through a vendor API. It had a public checkpoint that a community could point to together.

The first Qwen-7B release reported pretraining on more than 2.2 trillion tokens spanning Chinese, English, code, and other multilingual material. The base model learned general continuation and representations; the Chat version received additional instruction and alignment training. Chinese support was not simply a matter of including Chinese pages in the corpus. Tokenizer choices determined how many characters and phrases fit into a fixed context and how much each prompt cost in tokens. Teams accustomed to translated English tone, thin Chinese knowledge, or inefficient segmentation encountered those choices in every interaction.

Distribution mattered as much as model design. ModelScope provided an entry point well connected to Chinese networks and Alibaba’s ecosystem; Hugging Face placed the weights inside the tooling already used by a global developer community. The repository supplied examples for Transformers and later added paths for quantization and fine-tuning. Day one did not deliver every adapter at once, but it established a loop: weights were downloaded, failures appeared in issues, code was updated, and a recognizable user base remained in place for the next release.

The license drew a clear boundary around that loop. Research use was allowed. Under the terms at the time, commercial entities below 100 million monthly active users could use the model, while those at or above the threshold needed separate permission. This was open-weight distribution with conditional commercial terms, not unrestricted Apache-style open source. The threshold left room for most individuals, universities, and startups to experiment; for the largest platforms, the license itself became part of the adoption decision.

Publisher benchmarks covered Chinese and English understanding, code, and mathematics. They showed what the team intended to demonstrate, not a substitute for independent evaluation. A more durable shift appeared in how the model could be used: companies could bring a Chinese base model inside their networks, researchers could repeat experiments on the same checkpoint, and developers could report failures against a public version. First place on a table would age quickly. A continuing distribution relationship meant that later releases no longer arrived as isolated pieces of news.

Qwen would expand into more sizes, vision, audio, code, and mathematics, while its licensing evolved. Those events should not be poured backward into 3 August. What happened that day was specific enough: a pair of 7B checkpoints trained with Chinese and English at the center landed on two model hubs at once. From then on, “the next Qwen” was not only a cloud-product update. It became a public event that developers would wait for, download, and compare on their own machines.