近期，国产大模型密集上新，且越来越多的大模型厂商挑选开源。例如，9月8日，北京面壁智能科技股份有限公司（以下简称“面壁”）联合OpenBMB开源新一代MiniCPM5-2B。8月，阿里巴巴集团控股有限公司（以下简称“阿里”）旗舰模型Qwen3.8-Max首次开放权重；北京智谱华章科技股份有限公司（以下简称“智谱”）发布GLM-5.3；腾讯混元开源Hy4preview。
“从纯文本到原生多模态，国产大模型正将核心能力推向开源社区。”有行业分析师向《证券日报》记者表示，这一轮开源潮既是工艺自信的集中展示，更是国产厂商围绕开发者生态、国产算力和商业化路径的一次集体卡位。
从“比规模”到“比省钱”
过去衡量大模型的第一指标是参数量，如今这套逻辑正被瓦解。旗舰模型“大而不贵”：Qwen3.8-Max、腾讯混元Hy4等新一代旗舰模型集体使用稀疏MoE架构，总参数站上万亿级，单词元激活却控制在千亿上下，再叠加百万词元长上下文，以极低的激活成本换取旗舰能力。
在端侧维度，8月以来，阿里27B稠密模型、蚂蚁7.9B百灵、科大讯飞4B/1.7B星火接连开源，面壁最新发布的MiniCPM5-2B更在34项基准实测中平均得分超越4B档最优模型。智能密度正在取代参数规模，成为厂商较量的新战场。
竞争标尺切换的背后，是训练范式的深层转变。智谱GLM-5.3在基座不变的情况下，仅靠后训练便达成编程能力五成提高；腾讯混元Hy4首次让模型参与自身训练方法与数据策略的自动优化，推理吞吐提高超三成。大模型能力的增长曲线，正从堆算力堆参数转向拼工程拼方法，各家迭代节奏也因而从半年压缩到一两个月。
路线分化同样清晰，万亿级MoE旗舰承接大型集团复杂业务；10B级稠密与全模态基座主攻PC、终端硬件本地部署；轻量化模型则将办公智能体、企业自动化流程作为主战场。一个覆盖云端—终端—边缘的分层开源矩阵已然形成。
上述行业分析师表示，这一轮密集开源的深层信号在于，国产大模型正在经历一场去泡沫化的结构性调整。当头部厂商不再以参数量作为核心卖点，而是集体转向推理成本、端侧部署和工程化效率时，说明行业共识已经形成，真正决定胜负的不是谁的模型更大，而是谁能让AI以足够低的成本跑进真实业务情形。“接下来，能跑通付费闭环的厂商会迅速拉开身位，剩下的则会被压缩成纯基础设施供应商。”
生态之争定胜负
如果说“比省钱”是竞赛的表层，那么开源带来的产业重估才是深层变量。
开源与普惠定价正在相互强化，点燃“词元经济”。国家数据局最新数据表明，2026年6月，我国日均词元调用量激增至近175万亿，稳居全球第一。OpenRouter平台数据表明，开源模型处理的词元占比从1月的34%升至6月的65%。
与此对应，头部厂商定价集体下探：腾讯混元Hy4输入仅6元/百万词元，星火X2.5多语言版低至1.6元/百万词元，Qwen3.8-Flash更是压到1元/百万词元。“九成能力、一成售价”的性价比，扫清了智能体规模化商用最现实的成本障碍，模型能力正加速渗透千行百业。
竞争胜负手也随之从模型本身转移到“模型+工具+生态”的体系化能力。智谱同步上线编程工具ZCode，腾讯将Hy4接入CodeBuddy、元宝、ima机型矩阵，阿里以“云+模型+编程工具链”构筑闭环，面壁则以全栈开源的方式开放数据与训练配方，换取开发者和终端厂商的生态粘附。开源已从权重可下载的初级阶段，迈向可商用、可复现、可二次开发的深水区。
广州艾媒数聚信息咨询股份有限公司CEO张毅对《证券日报》记者表示，落地深水区并非坦途。一维度，开源基座快速增多抬高了企业选型成本，各家商用授权条款并不统一，部分旗舰基座采取收入分成模式，商业化路径仍待打磨；另一维度，算力与流量成本的波动同样可能扰动落地节奏。
“以8月下旬轻量化模型集中开源，9月初科大讯飞、面壁接续上新为标志，国内开源大模型已全面进入工程化落地周期。”张毅表示，随着基座间表现差距不断收窄，谁能依托开源基座在办公、工业、终端情形跑通可复制的商业化案例，谁就能在下一轮行业位次重排中占得先机。
本文来源：https://www.stcn.com/article/detail/4177402.html
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Like
Share
Recently, Chinese domestic large models have seen a dense wave of new releases, with more and more vendors opting to open-source their work. For instance, on September 8, MiniMax (hereinafter referred to as "MiniMax") jointly open-sourced the new MiniCPM5-2B with OpenBMB. In August, Alibaba Group's flagship model Qwen3.8-Max opened its weights for the first time; Zhipu AI released GLM-5.3; and Tencent Hunyuan open-sourced Hy4preview.
"From pure text to native multimodal, Chinese domestic large models are pushing their core capabilities into the open-source community," an industry analyst told *Securities Daily*. "This wave of open-sourcing is both a concentrated showcase of technical confidence and a collective strategic move by domestic vendors around developer ecosystems, domestic computing power, and commercialization pathways."
From "Comparing Scale" to "Comparing Cost Savings"
In the past, the first metric for evaluating large models was parameter count, but this logic is now being upended. Flagship models are becoming "large yet affordable": next-generation flagships like Qwen3.8-Max and Tencent Hunyuan Hy4 all adopt sparse MoE architecture, with total parameters exceeding a trillion while per-token activated parameters stay around hundreds of billions. Coupled with million-token long context windows, this delivers flagship-level capability at extremely low activation costs.
On the edge-device front, since August, Alibaba has open-sourced its 27B dense model, Ant Group its 7.9B Bailing, and iFlytek its 4B/1.7B Spark series. MiniMax's latest MiniCPM5-2B even outperforms top 4B-tier models across 34 benchmark tests on average. Intelligence density is replacing parameter scale as the new battleground for vendors.
Behind the shift in competitive yardsticks lies a deeper transformation in training paradigms. Zhipu's GLM-5.3 achieved a 50% improvement in programming capability through post-training alone, without changing its base model. Tencent Hunyuan Hy4 allowed the model to participate in optimizing its own training methods and data strategies for the first time, boosting inference throughput by over 30%. The growth curve of large model capabilities is shifting from stacking compute and parameters toward competing on engineering and methodology, compressing iteration cycles from six months down to one or two months.
Route differentiation is equally clear: trillion-parameter MoE flagships serve complex enterprise workloads, 10B dense and fully multimodal bases target local deployment on PCs and终端 hardware, and lightweight models focus on office agents and enterprise automation. A layered open-source matrix spanning cloud—endpoint—edge has taken shape.
The analyst noted that the deeper signal behind this wave of dense open-sourcing is that domestic large models are undergoing a structural deflation adjustment. When leading vendors no longer pitch parameter counts as their core selling point but collectively pivot toward inference cost, edge deployment, and engineering efficiency, it signals industry consensus: the real decider of victory is not whose model is bigger, but whose AI can run into real business scenarios at sufficiently low cost. "Vendors that can establish a viable paid loop will soon pull ahead; the rest will be compressed into pure infrastructure suppliers."
Ecosystem Competition Decides the Winner
If "comparing cost savings" is the surface level of the competition, then the industrial revaluation driven by open source is the deeper variable.
Open source and inclusive pricing are mutually reinforcing, igniting the "token economy." Latest data from the National Data Administration shows that in June 2026, China's daily token call volume surged to nearly 175 trillion, firmly holding the global No. 1 position. Platform data from OpenRouter shows that the share of tokens processed by open-source models rose from 34% in January to 65% in June.
In response, pricing from leading vendors has collectively dropped: Tencent Hunyuan Hy4 input is priced at just 6 yuan per million tokens, Spark X2.5 multilingual edition as low as 1.6 yuan per million tokens, and Qwen3.8-Flash pushed down to 1 yuan per million tokens. This cost-effectiveness of "90% capability at 10% price" has cleared the most realistic cost barrier to scaled commercial adoption of AI agents, accelerating model capability penetration across industries.
The competitive winning hand has therefore shifted from the model itself to systematic capabilities of "model + tools + ecosystem." Zhipu launched its programming tool ZCode in tandem; Tencent integrated Hy4 into its CodeBuddy, Yuanbao, and ima product matrices; Alibaba built a closed loop with "cloud + model + programming toolchain"; and MiniMax open-sourced its data and training recipes via a full-stack approach to secure developer and终端 vendor ecosystem stickiness. Open source has moved beyond the初级 stage of downloadable weights into the deeper waters of commercial viability, reproducibility, and secondary development.
Zhang Yi, CEO of Guangzhou iiMedia Data聚合Information Consulting Co., Ltd., told *Securities Daily* that entering the deep water zone is no easy road. On one hand, the rapid proliferation of open-source bases raises enterprises' selection costs, with inconsistent commercial licensing terms across vendors and some flagship bases adopting revenue-sharing models—leaving commercialization pathways still needing refinement. On the other hand, volatility in compute and traffic costs could also disrupt deployment timelines.
Zhang Yi stated, "Marked by the concentrated open-sourcing of lightweight models in late August and successive new releases from iFlytek and MiniMax in early September, China's open-source large model ecosystem has fully entered its engineering deployment phase." With performance gaps between bases narrowing, whoever can build replicable commercial use cases on top of open-source bases in office, industrial, and终端 scenarios will seize the initiative in the next round of industry reshuffling.
Editor-in-charge: Deng Weiping
MiniMax
Alibaba
Zhipu AI
Recommendations for You
MiniMax Rockchip
MiniMax Beijing
Source: https://www.stcn.com/article/detail/4177402.html. For copyright concerns, please email: 13951431913#139.com or submit a comment via the panel on the right.