讯飞星火发布

2023 年 5 月 6 日，科大讯飞在合肥发布讯飞星火认知大模型。对很多人来说，这可能是第一次认真看这家公司——它以语音识别和翻译起家，产品覆盖教育硬件、办公、车载语音，是典型的「垂直场景巨头」。当中国大模型竞争在文心一言之后迅速升温时，讯飞选择下场，而且把发布会开得声势浩大。

发布会上，讯飞展示了星火的文本生成、知识问答、数学推理、代码能力和多语种翻译，还现场演示了与虚拟人结合的多轮互动。讯飞的思路很清晰：不做「又一个聊天机器人」，而是把大模型能力接进自己已有的教育、办公、车载、客服场景。现场公布的计划里，星火将进入学习机、办公本、智能座舱等讯飞硬件——这是纯互联网公司没有的落地渠道。

星火的底气来自讯飞多年的积累。语音识别、自然语言理解、机器翻译这些底层技术，讯飞已经做了二十年，中文语料与场景数据是它的核心资产。发布会上讯飞公布了后续迭代路线图，承诺按季度升级。这种「发布会就画路线图」的做法，后来成为中国大模型厂商的标准操作——先用承诺锁定预期，再逐步兑现。

星火发布的意义，在于它把「百模大战」的参与者结构变复杂了。此前入场的主要是互联网大厂和 AI 创业公司；讯飞代表的，是拥有垂直场景、硬件渠道和行业数据的一类玩家。它们不追求通用对话的全面领先，而是用「场景+模型」的组合建立护城河。这种路径，后来在医疗、教育、办公等领域被反复复制。

回看 2023 年 5 月，星火发布会没有改变通用大模型的排名格局，但它展示了另一种打法：不抢第一，但保证「我的场景里我最懂」。讯飞后来把星火深度整合进学习机和办公产品，让数千万用户在真实场景里用上了国产大模型。当行业争论「谁能赢」时，这类垂直巨头用一条不同的路线证明了自己的一席之地。

On May 6, 2023 iFlytek released the Spark (Xinghuo) cognitive large model in Hefei. For many, it was the first serious look at this company—known for speech recognition and translation, with products spanning education hardware, office tools, and in-car voice. It is a classic vertical-scenario giant. When China's large-model race heated up after Ernie Bot, iFlytek chose to join, and did so with a loud event.

At the event iFlytek showed Spark's text generation, knowledge QA, math reasoning, coding, and multilingual translation, plus a live multi-turn interaction with a virtual human. The strategy was clear: not another chatbot, but a large model wired into iFlytek's existing education, office, automotive, and customer-service scenarios. The announced plans put Spark into learning devices, office notepads, and smart cockpits—distribution channels pure internet companies lack.

Spark's confidence came from years of accumulation. Speech recognition, natural-language understanding, and machine translation are technologies iFlytek has worked on for two decades; Chinese-language corpora and scenario data are its core assets. The event also published a quarterly upgrade roadmap—a practice that later became standard for Chinese model vendors: lock expectations with promises first, then deliver.

Spark's release mattered because it complicated the roster of the hundred-models war. Before it, entrants were mainly internet giants and AI startups. iFlytek represented a class of player with vertical scenarios, hardware channels, and industry data. These companies did not pursue all-around leadership in general dialogue; they built moats with "scenario plus model" combinations—a path later copied across healthcare, education, and office work.

Looking back at May 2023, the Spark event did not change the ranking of general-purpose models, but it demonstrated another play: do not aim for first, but guarantee "I know my scenarios best." iFlytek later integrated Spark deeply into learning devices and office products, putting a domestic large model into tens of millions of users' real scenes. While the industry argued over who would win, this class of vertical giant proved its place with a different route.