Specialization Beats Scale: A Strategic Variable Most AI Procurement Decisions Overlook
Specialization Beats Scale: A Strategic Variable Most AI Procurement Decisions Overlook
专精胜过规模：大多数 AI 采购决策所忽视的战略变量
When a model’s training history is moved close enough to its deployment task, parameter count stops being the decisive variable. A 3-billion-parameter specialized model outperformed every commercial frontier API tested in a well-measured enterprise domain — at roughly fifty times lower cost. 当模型的训练历史与其实际部署任务足够接近时，参数量就不再是决定性变量。在一个经过严谨评估的企业领域中，一个 30 亿参数的专精模型击败了所有参与测试的商业前沿 API，且成本降低了约 50 倍。
In April, we released DharmaOCR — a pair of specialized small language models for structured OCR, alongside a benchmark and the accompanying paper. The models and the benchmark are available on Hugging Face. Together they form part of a broader effort at Dharma to study how specialization, alignment, and inference economics interact in production AI systems. This article isolates one strategic implication from those findings: the relationship between specialization, distributional alignment, and parameter scale. What follows develops it within the boundaries the paper supports. 今年 4 月，我们发布了 DharmaOCR——一对用于结构化 OCR 的专精型小语言模型，并附带了基准测试和相关论文。这些模型和基准测试已在 Hugging Face 上发布。它们共同构成了 Dharma 更广泛研究计划的一部分，旨在探讨专精化、对齐和推理经济学如何在生产级 AI 系统中相互作用。本文将从这些研究结果中提炼出一个战略启示：专精化、分布对齐与参数规模之间的关系。下文将在论文所支持的范围内对此进行阐述。
For the past three years, enterprise AI strategy has largely operated on a stable assumption: the safest choice was usually the largest frontier model available. Smaller models were considered primarily where the workload could tolerate some reduction in quality in exchange for lower cost. The logic behind that assumption was straightforward. Capability appeared to scale with parameter count, frontier providers consistently led the major benchmarks, and the cost of choosing the wrong model was often perceived as greater than the cost of paying for the leading one. 在过去三年中，企业 AI 战略在很大程度上基于一个稳定的假设：最稳妥的选择通常是市面上最大的前沿模型。小型模型主要在工作负载可以容忍一定质量下降以换取更低成本的情况下才会被考虑。这一假设背后的逻辑很简单：能力似乎随参数量增加而扩展，前沿模型提供商始终在各大基准测试中领先，且选择错误模型的代价通常被认为高于支付领先模型费用的代价。
The reasoning is defensible. But the empirical record now includes a result that the comparison set behind it cannot easily explain. Earlier this year, Dharma published a benchmark in which a 3-billion-parameter model — specialized through a fine-tuning pipeline any well-resourced enterprise could replicate — outperformed every commercial frontier API tested. Not by a small margin, and not on a metric a buyer would dismiss. The cost gap ran in the opposite direction from the quality gap: the highest-scoring model was also the cheapest to operate, by a margin large enough to alter procurement arithmetic at any meaningful volume. 这种推理是有道理的。但目前的实证记录中出现了一个结果，是其背后的比较集无法轻易解释的。今年早些时候，Dharma 发布了一项基准测试，其中一个 30 亿参数的模型——通过任何资源充足的企业都能复制的微调流程进行专精化——击败了所有参与测试的商业前沿 API。这并非微弱优势，也不是买家会忽视的指标。成本差距与质量差距呈现相反方向：得分最高的模型同时也是运行成本最低的模型，其差距足以改变任何有意义规模下的采购算账方式。
The result is not isolated. It is the most rigorously measured instance, to date, of a pattern Dharma has observed across other domains — and one a growing body of specialization research has begun to document (Subramanian et al., 2025; Pecher et al., 2026). But it does raise a question worth asking explicitly: when the largest model is not the best-performing model, what variable is doing the work? 这一结果并非孤例。这是迄今为止 Dharma 在其他领域观察到的模式中最严谨的测量实例，也是越来越多的专精化研究开始记录的现象（Subramanian 等人，2025；Pecher 等人，2026）。但这确实提出了一个值得明确探讨的问题：当最大的模型不是表现最好的模型时，究竟是什么变量在起作用？
The Strategic Default / 战略默认项
The procurement default did not arrive by accident. It arrived because, for most of the past three years, it was correct. When GPT-4 was released, it outperformed every smaller model on the benchmarks that mattered. The pattern repeated, with refinements, through Claude 3, Gemini 1.5, and each generation of frontier release in 2025. Capability scaled with parameter count and with training compute (Kaplan et al., 2020) — the empirical relationship OpenAI’s scaling laws had formalized years earlier. The lesson followed: a buyer who picked the largest model available was, on average, picking the best-performing tool. 这种采购默认项并非偶然出现。它的出现是因为在过去三年的大部分时间里，它是正确的。当 GPT-4 发布时，它在所有重要的基准测试中都击败了每一个小型模型。这种模式在 Claude 3、Gemini 1.5 以及 2025 年每一代前沿模型发布中不断重复并得到完善。能力随着参数量和训练算力（Kaplan 等人，2020）的增加而扩展——这是 OpenAI 的缩放定律（Scaling Laws）多年前就已形式化的实证关系。随之而来的经验是：选择市面上最大模型的买家，平均而言，就是在选择表现最好的工具。
The assumption was defensible because, for most of the comparisons that produced it, it was correct. What changed was not that the assumption had always been wrong. What changed was that the comparison set on which it rested may not have been complete. What was missing was a different kind of model. Not a smaller frontier model. A specialized model — one whose training history had been deliberately moved closer to the task it would be asked to do, through a sequence of fine-tuning steps that adapted a smaller base to the domain it would be deployed in. The paper described in the opening is among the first to run that comparison with cost, quality, and production stability measured side by side. 这一假设之所以站得住脚，是因为在产生该假设的大多数比较中，它是正确的。改变的并不是假设一直以来都是错的，而是它所依赖的比较集可能并不完整。缺失的是另一种模型。不是更小的前沿模型，而是专精模型——通过一系列微调步骤，将较小的基础模型适配到其部署领域，从而使其训练历史被刻意地向其任务需求靠拢。开篇提到的论文是首批将成本、质量和生产稳定性进行并排比较的研究之一。
What the Empirical Record Actually Shows / 实证记录究竟显示了什么
The benchmark used in the paper was a domain-specific evaluation: Brazilian Portuguese OCR across printed documents, handwritten text, and legal and administrative records. The benchmark itself is not the point of this article. What matters is what it measured, and the comparisons it ran. On extraction quality, the highest-scoring model in the comparison was the specialized 3-billion-parameter model. It scored 0.911 on the benchmark’s composite score, which combines edit-distance similarity with n-gram overlap. 论文中使用的基准测试是一项特定领域的评估：针对巴西葡萄牙语的 OCR，涵盖印刷文档、手写文本以及法律和行政记录。基准测试本身并非本文重点，重要的是它测量的内容及其进行的比较。在提取质量方面，比较中得分最高的模型是那个 30 亿参数的专精模型。它在基准测试的综合得分（结合了编辑距离相似度和 n-gram 重叠度）中获得了 0.911 分。
The closest frontier alternative — Claude Opus 4.6 — scored 0.833. Below it: Gemini 3.1 Pro at 0.820, GPT-5.4 at 0.750, Google Vision at 0.686, Google Document AI at 0.640, GPT-4o at 0.635, Amazon Textract at 0.618, and Mistral OCR 3 at 0.574. The specialized model finished first, and the gap to Claude Opus 4.6 — close to eight percentage points — was wider than any other gap between adjacent finishers in the comparison. 最接近的前沿替代方案 Claude Opus 4.6 得分为 0.833。其后依次为：Gemini 3.1 Pro（0.820）、GPT-5.4（0.750）、Google Vision（0.686）、Google Document AI（0.640）、GPT-4o（0.635）、Amazon Textract（0.618）和 Mistral OCR 3（0.574）。专精模型位居第一，且与 Claude Opus 4.6 之间近 8 个百分点的差距，比比较中任何其他相邻名次之间的差距都要大。
On cost, the gap was far wider. The specialized 3B model ran at approximately fifty-two times lower cost per million pages than Claude Opus 4.6 — a margin computed from inference-infrastructure cost against published API pricing. The quality–cost picture, plotted as a Pareto frontier, shows the specialized model in the upper-left of the chart, with the commercial APIs below and to the right. 在成本方面，差距则大得多。该 30 亿参数专精模型的每百万页运行成本比 Claude Opus 4.6 低约 52 倍——这一差额是根据推理基础设施成本与公开 API 定价计算得出的。以帕累托前沿（Pareto frontier）绘制的质量-成本图显示，专精模型位于图表的左上角，而商业 API 则位于右下方。