基座没变。能力涨了50%。
8月14日，智谱正式发布GLM-5.3。最让人意外的不是它有多强，而是它变强的方式：基座模型完全不变，所有提升都来自后训练。
换句话说，同样的参数规模、同样的预训练模型，只是通过后训练阶段的Scaling——更大规模的合成数据、更长的强化学习时间、更丰富的任务环境——就把编程能力提升了约50%。
这不是一次常规的版本迭代。这是一个信号：Scaling Law的下半场，可能不是堆参数，而是拼后训练。
后训练Scaling的威力
先看一组具体数据。GLM-5.2到GLM-5.3，基座一个参数没动，但：
Terminal Bench 3.0从4.6分涨到28.3分；DeepSWE v1.1从46.2分涨到66.9分；内部Z.ai Code Bench提升约50%；Agents' Last Exam同样取得开源SOTA。
这些提升全部集中在编程和Agent任务上。通用推理能力提升不大，但在需要多步骤、需要工具调用、需要长时间执行的任务上，进步是飞跃式的。
为什么后训练能带来这么大的提升？因为预训练教会模型的是"知识"——世界是什么样的、语言怎么组织。而后训练教会模型的是"能力"——怎么思考、怎么解决问题、怎么跟环境交互。
"以前大家拼的是谁的基座更大。现在发现，基座可能只是地基——真正决定房子能盖多高的，是后训练这层结构。"—— 一位AI技术研究者
这个发现对行业的影响是深远的。过去几年，整个行业的叙事都是"越大越好"——参数越多、算力越多、数据越丰富，模型就越强。Scaling Law似乎是铁律。
但GLM-5.3的故事告诉我们：基座的大小可能不是唯一的瓶颈。当基座达到一定规模之后，后训练的质量和规模，可能比继续堆参数更有效率。
这对于算力相对有限的团队来说，尤其重要——你不需要训练最大的模型，也能通过极致的后训练，在特定领域做到最强。
开源编程王座易主
GLM-5.3还有一个重要承诺：两周内完成安全加固后开源权重。
这意味着什么？意味着开源社区即将迎来一个编程能力最强的基础模型。在此之前，开源编程的最强者可能是Qwen或者DeepSeek，但GLM-5.3在编程基准上的表现，已经超过了它们。
更重要的是，这个模型是完全开源的——不是"开源可商用"那种有各种限制的半开源，而是真正可以自由使用、修改、分发的开源权重。对于开发者和创业公司来说，这相当于免费拿到了一个顶级的编程模型。
而且，因为GLM-5.3的提升主要来自后训练，这也给开源社区指明了一个方向：拿到一个好的基座之后，你可以自己炼后训练。不需要千亿算力，只要方法对路，用相对少的资源也能大幅提升模型在特定领域的能力。
这可能会催生一波基于开源基座的垂直领域模型——医疗、法律、金融、教育——每个领域都可以用自己的数据做后训练，做出垂直领域的专家模型。
当然，GLM-5.3也不是完美的。它的优势集中在编程和Agent任务上，通用推理能力跟顶级闭源模型还有差距。但对于编程开发者来说，一个开源的、顶级的、可以本地部署的编程模型——这个诱惑力已经足够大了。
后训练Scaling的时代，可能才刚刚开始。
明天见。
Same base. 50% more capable.
On August 14, Zhipu officially launched GLM-5.3. What's most surprising isn't how strong it is - it's how it got stronger: the base model is completely unchanged. All improvements come from post-training.
In other words, same parameter scale, same pre-trained model - just through post-training scaling - larger-scale synthetic data, longer reinforcement learning time, richer task environments - and coding capability improved by roughly 50%.
This isn't a routine version update. It's a signal: the second half of Scaling Law might not be about stacking more parameters - it's about better post-training.
The Power of Post-Training Scaling
Let's look at some concrete numbers. From GLM-5.2 to GLM-5.3, not a single parameter in the base changed, yet:
Terminal Bench 3.0 went from 4.6 to 28.3; DeepSWE v1.1 went from 46.2 to 66.9; internal Z.ai Code Bench improved by ~50%; Agents' Last Exam also achieved open-source SOTA.
These improvements are all concentrated in coding and agent tasks. General reasoning capability didn't improve much, but on tasks requiring multi-step reasoning, tool use, and extended execution - the leap is dramatic.
Why can post-training bring such big improvements? Because pre-training teaches a model "knowledge" - what the world is like, how language works. Post-training teaches a model "capability" - how to think, how to solve problems, how to interact with the environment.
"Everyone used to compete on who had the bigger base model. Now we're discovering the base might just be the foundation - what really determines how tall you can build is the post-training structure." - An AI Researcher
This discovery has profound implications for the industry. For the past few years, the entire industry's narrative has been "bigger is better" - more parameters, more compute, more data equals stronger models. Scaling Law seemed like an iron law.
But GLM-5.3's story tells us: base model size might not be the only bottleneck. Once a base reaches a certain scale, post-training quality and scale might be more efficient than stacking more parameters.
This is especially important for teams with relatively limited compute - you don't need to train the biggest model to be the best in a specific domain, as long as you nail post-training.
Open-Source Coding Crown Changes Hands
GLM-5.3 also comes with an important promise: open-source weights after security hardening within two weeks.
What does this mean? It means the open-source community is about to get its strongest-ever coding foundation model. Before this, the open-source coding leader was probably Qwen or DeepSeek - but GLM-5.3's performance on coding benchmarks has already surpassed them.
More importantly, this model is fully open source - not "open-source with restrictions" semi-open, but truly free to use, modify, and distribute. For developers and startups, that's like getting a top-tier coding model for free.
And because GLM-5.3's improvements mainly come from post-training, it also points the open-source community in a direction: once you have a good base, you can do your own post-training. You don't need billions in compute - with the right approach, you can dramatically improve a model's capability in specific domains with relatively few resources.
This could spawn a wave of vertical domain models built on open-source bases - healthcare, law, finance, education - each domain using its own data for post-training to create expert models.
Of course, GLM-5.3 isn't perfect. Its advantages are concentrated in coding and agent tasks; general reasoning still lags behind top closed-source models. But for coding developers, an open-source, top-tier, locally-deployable coding model - that's already plenty of appeal.
The era of post-training scaling might have just begun.
See you tomorrow.
以前大家拼的是谁的基座更大。现在发现，基座可能只是地基——真正决定房子能盖多高的，是后训练这层结构。
—— 一位AI技术研究者
Everyone used to compete on who had the bigger base model. Now we're discovering the base might just be the foundation - what really determines how tall you can build is the post-training structure.
- An AI Researcher
