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.
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.
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.
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.
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?
The Strategic Default
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. In the absence of a more discriminating signal, defaulting to scale was the rational move.
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.
What the Empirical Record Actually Shows
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. 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.
Results of the models evaluated on DharmaOCR-Benchmark. Parentheses in the first column indicate the specialization techniques used. When a model is not indicated as LoRA, it means that full fine-tuning has been performed. Entries marked with “Quant” indicate AWQ-quantized variant with best performance among the quantized configurations.
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. (The financial-modeling depth is developed in The Real Economics of Text Degeneration.)
On production stability, the same model produced the lowest text-degeneration rate evaluated — a measure of how often a generation enters a self-reinforcing loop and fails to produce a usable output. (The production-stability case is developed in the cluster’s Text Degeneration article.) The 3B model recorded 0.20% on this benchmark; the next closest specialized model, 0.40%; the larger general-purpose open-source baselines ran higher; the commercial APIs were not benchmarked on this metric directly.
Text degeneration rate (%) across alignment stages. SFT reduces degeneration relative to vanilla models in most cases, whereas DPO further reduces it, even compared to the SFT-tuned model.
These three findings — quality, cost, and stability, all led by the same 3B specialized model — are the article’s empirical anchor. Together, they make the empirical case stronger than any single finding would alone. The paper does not claim, and this article does not claim, that the result generalizes to every enterprise AI workload. What it claims is that on this benchmark, the smallest specialized model in the experiment was first on every dimension that mattered.
Which makes the obvious question the right question. The smallest model in the comparison won on quality, on cost, and on stability. Parameter count, by itself, does not explain that result. The natural follow-up — identifying the variable that does — is where the conversation moves next.
The Variable That Mattered
Part of this is intuitive. A 3-billion-parameter model focused on the deployment task will often outperform a much larger model whose parameters are spread across material the task will never touch — other languages, other corpora, other domains. What the paper adds goes further: one of the important variables is not only how parameters are allocated, but how the model’s training history has been moved toward the task. In the experiments reported, this variable predicted relative performance more reliably than any other tested — including parameter count.
The paper names this directly. In its discussion, the authors describe the result as supporting the claim that “contextual specialization can be more decisive than number of model parameters alone.” What determined whether a model performed best was not parameter count, but how close its training trajectory had been moved to its deployment task. A larger model trained on a wider distribution finished below a smaller model trained on a narrower one. The narrower training was the variable that produced the win.
This is a different way of thinking about model performance than the procurement default invites. Under the default, parameter count is the dominant variable and training history is a secondary modifier. Under the framing the paper proposes, the priority reverses. Distributional alignment to the task becomes the dominant variable. Parameter count becomes one factor among several that shape how much benefit a given alignment step produces.
Specialization is not a way to compensate for being small. It is a way to be aligned.
The numbers bear the framing out. The 3B Nanonets-OCR2 — already specialized for general OCR before the paper began — was fine-tuned on the target domain through supervised fine-tuning and Direct Preference Optimization, and reached 0.921 with a 0.20% degeneration rate. A 3B general-purpose model of identical architecture, Qwen2.5-VL-3B, was run through the same procedure and reached 0.793 with 1.41% degeneration. Same architecture, same training, different result. The variable was the distance the model had already traveled toward the task before the procedure began.
Distributional alignment, on the framing the paper proposes, is not specific to OCR. It is a property of the relationship between a model and the task it is asked to perform. The question of which model is best for a given enterprise workload is, on this framing, mostly a question of how aligned its training history is — not how large the model is.
If distributional alignment is one of the variables that mattered most, the next question is how it accumulates. The paper’s evidence suggests it does not arrive in a single step. The result above turns out to be one instance of a broader pattern: specialization, in the paper’s data, behaves less like a binary state than like a hierarchy through which a model can be moved one step at a time.
Specialization Compounds
Alignment is not a single thing a model either has or lacks. It is a position on a hierarchy that can be moved up one step at a time. A general-purpose model sits at the bottom; a general-domain specialist (trained for the broader category of work) sits above it; a domain specialist (trained for the specific work it will be deployed on) sits above that. The same downstream training produces different results depending on which step the model starts from.
The paper’s evidence for this is structural. Two pairs of comparisons illustrate it directly.
At the 7-billion-parameter scale: the best fine-tuned model derived from Qwen2.5-VL-7B-Instruct — a general-purpose start — reached 0.906 with a 1.01% degeneration rate. The same training applied to olmOCR-2–7B — already specialized for general OCR — reached 0.927 with 0.40% degeneration. The quality gain was approximately 2.3 percent; the degeneration rate fell by nearly half. Same architecture, same data, same training pipeline. The variable was the starting position.
At the 3-billion-parameter scale (the comparison introduced earlier): Qwen2.5-VL-3B finished at 0.793 with 1.41% degeneration; Nanonets-OCR2–3B finished at 0.921 with 0.20% degeneration. Same procedure, same architecture class, different starting position. The quality gain was approximately 16 percent; the degeneration rate fell by a factor of roughly seven.
Progressive specialization strategy and comparison of two training paths. Three specialization levels are shown — vanilla generalist (Level 1), general-domain OCR specialist (Level 2), and domain-specific OCR specialist (Level 3) — plus a projected Level N for future sub-domain specialization.
Two pairs, two parameter scales, two consistent results. Specialization accumulates. A model already moved closer to the broader category of its eventual task benefits more from the same domain-specific training than a model starting from a wider distribution. The procedure does not produce alignment from nothing. It builds on whatever alignment is already present.
There are levels of specialization, and each level builds on the distribution encoded by the one before it. Multiple stages of training can progressively move a model closer to the target task distribution, producing materially different downstream outcomes even under similar architectural and computational constraints.
That pattern — alignment as an accumulating quantity — is the article’s strongest claim from the paper’s evidence. Its boundaries deserve to be marked explicitly. The hierarchy was demonstrated in one domain, on one benchmark, with two pairs of model comparisons. The mechanism has no domain-specific reason to be confined to OCR — but the evidence has not yet been gathered elsewhere, and an argument that respects its boundaries should mark that distinction. Expanding that empirical investigation across additional enterprise domains is part of the broader research direction this work opens, and that Dharma intends to investigate further across additional enterprise domains.
With that boundary marked, the strategic conversation moves forward. A variable shown to dominate parameter count in one well-measured enterprise domain is one strategy teams now have reason to weigh — not in every setting, but in any where the alignment test can be run.
The Strategic Questions That Change
A useful way to read the paper is not as an instruction for what enterprises should do next, but as a prompt for what they should ask. Three questions come into sharper focus.
The first: whether distributional alignment should be elevated alongside parameter count as a first-class variable in serious AI evaluation. The paper’s evidence does not argue for elevating it above parameter count. It argues, more modestly, that alignment is large enough as a variable to be tested explicitly rather than assumed to be small.
The second follows: is benchmark leadership, on its own, sufficient evidence for an enterprise procurement decision? In one well-measured domain, the model that led the public benchmarks was not the model that delivered the best result. If that divergence appears in other domains — and the paper does not establish that it does, only that it can — enterprise evaluation may need an additional layer of evidence, run on workloads representative of the deployment.
The third is about architecture, not method. If alignment is a position on a hierarchy that compounds, the choice of starting model — not only the fine-tuning procedure — becomes a strategic decision in its own right. A starting model already closer to the deployment task may produce materially better outcomes than a larger, more general model under the same training budget. But the deeper implication may be organizational rather than procedural. If specialization compounds, enterprises may eventually benefit less from searching for a single universally capable model than from building an ecosystem of models progressively aligned to their own domains, workflows, and operational constraints. Whether that architecture proves advantageous in practice is a question each organization has to evaluate within its own environment.
A Bounded Reframe
The article’s contribution is narrow, by design. It has not argued that frontier models are inferior, or disposable, or that the procurement default should be inverted. It has argued, on the strength of one paper’s evidence, that frontier models are not necessarily the best-performing choice for every enterprise AI workload. In the experiments reported, smaller specialized models with training histories more closely aligned to the deployment task achieved superior quality, lower cost, and greater production stability than the larger commercial APIs evaluated. The implication is not that frontier models are inferior. It is that specialization history may be a more strategically important variable for enterprise AI systems than many evaluation frameworks currently assume.
We wrote this article not to argue that scale no longer matters, but to isolate a variable the current enterprise AI conversation may still underweight. Training history can be observed, evaluated, and moved closer to a deployment task through successive stages of specialization. In the comparisons reported in the paper, that relationship materially changed the ranking of every model evaluated. Whether it changes rankings elsewhere is a question for the next set of experiments.
Sources:
Cardoso, Gabriel Pimenta de Freitas, et al. “DharmaOCR: Specialized Small Language Models for Structured OCR that outperform Open-Source and Commercial Baselines.” arXiv preprint arXiv:2604.14314 (2026).
Subramanian, Karthik, et al. “Small Language Models for Domain-Specific Enterprise AI Workloads.” arXiv preprint arXiv:2503.11872 (2025).
Pecher, Branislav, et al. “Comparing specialised small and general large language models on text classification: 100 labelled samples to achieve break-even performance”, (2026).
Kaplan, Jared, et al. “Scaling Laws for Neural Language Models.” arXiv preprint arXiv:2001.08361 (2020).