TL;DR. A 3B model specialised on Brazilian Portuguese OCR outscores Claude Opus 4.6 — 0.911 versus 0.833, per Dharma-AI — at 52 times lower cost per million pages. Nemotron-Labs Diffusion reaches 6.4× the throughput of a standard autoregressive model on B200 hardware, per NVIDIA. Three model categories. Three distinct selection criteria: domain fit, cost, and throughput.
Three years of procurement defaults — and why they are breaking
Since 2023, the dominant heuristic in enterprise AI procurement has stabilised around a single principle: the largest available model is the safest choice. The reasoning was defensible — frontier models absorbed edge cases, avoided the blind spots of premature specialisation, and externalised maintenance risk.
Two technical publications, appearing three days apart on Hugging Face, shift that frame. On 22 May 2026, Dharma-AI published a comparative benchmark on a corpus of Brazilian Portuguese legal and administrative OCR documents, pitting a 3-billion-parameter specialised model against the leading frontier models. On 23 May, NVIDIA published the Nemotron-Labs Diffusion family, introducing a block-based generation mode that reaches 6.4× the speed of a standard autoregressive baseline. Both publications share a common subtext: model size is not the only axis of enterprise competitiveness. Two others now demand measurement — distributional alignment to the deployment task, and inference throughput.
Where specialised models take the lead
On the Dharma-AI benchmark — covering printed, handwritten, and administrative documents in Brazilian Portuguese — the Dharma-OCR 3B model scores 0.911. Claude Opus 4.6 reaches 0.833, Gemini 3.1 Pro 0.820, GPT-5.4 0.750, GPT-4o 0.635, and Amazon Textract 0.618, per the Dharma-AI publication. The gap between first and second place is 7.8 percentage points.
Cost is the decisive argument at scale. Dharma-OCR 3B costs 52 times less than Claude Opus 4.6 per million pages processed, according to the same source.
Production stability is the third differentiator. On text degeneration rate — a critical metric in automated pipelines where models produce incoherent or repetitive output — Nanonets-OCR2 3B records 0.20%, against 1.41% for Qwen2.5-VL-3B in general-purpose use, per Dharma-AI. The ratio is 7 to 1. olmOCR-2 7B, another OCR specialist, reaches 0.40% — well below the general-purpose model of comparable size.
The structural logic behind these results is made explicit by Dharma-AI: specialisation compounds across levels. At 7 billion parameters, moving from a general-purpose model to a generic OCR specialist improves quality by 2.3% and halves the degeneration rate. At 3 billion parameters, the quality gain reaches 16% and the degeneration rate drops by a factor of seven, per the same publication.
Where frontier and diffusion models hold their ground
Frontier models: versatility as structural advantage
The Dharma-AI article is explicit on scope: the results cover a single, well-measured domain. On multi-domain tasks, complex reasoning over variable perimeters, or use cases whose boundaries are undefined at procurement time, frontier models retain an operational advantage that specialists cannot replicate. A model scoring 0.833 on Portuguese OCR may score 0.95 on a different domain — or be the only model capable of handling an unforeseen request type. Dharma-AI does not argue that frontier models are obsolete; the argument is that their dominance is not universal.
Nemotron-Labs Diffusion: throughput as infrastructure differentiator
The Nemotron-Labs family — 3B, 8B, 14B — introduces three distinct generation modes, per NVIDIA. Standard autoregressive mode. Block-based diffusion mode, generating 2.6× more tokens per forward pass. Self-speculation mode, which uses diffusion as a draft and autoregressive verification as a final check, reaching 6.4× baseline speed and approximately 865 tokens per second on B200 hardware, per the NVIDIA publication.
The critical technical point: this throughput gain is lossless at temperature zero. The output is identical to autoregressive mode — not an approximation. Nemotron-Labs Diffusion 8B also shows 1.2% higher average accuracy than Qwen3 8B, per the same source. On general reasoning benchmarks, frontier models retain their advantage — Nemotron-Labs Diffusion is positioned as an inference engine for latency- and throughput-constrained workloads, not as a frontier challenger.
Pricing and operational implications
Three cost and infrastructure profiles emerge, without the categories being mutually exclusive:
- Specialised models: very low marginal cost per request (52× documented cost reduction on OCR, per Dharma-AI). Upfront cost: domain data annotation, fine-tuning, validation. Break-even depends on the volume of homogeneous requests and the organisation's annotation cost.
- Frontier models via API: no proprietary infrastructure, no fine-tuning. Usage-based billing. High cost at scale, but maintenance and updates externalised. Relevant for low-frequency tasks or variable-scope use cases.
- On-premises diffusion models: a 6.4× throughput gain frees inference slots on existing infrastructure, per NVIDIA. The critical variable is hardware compatibility — the self-speculation mode is documented on B200 — and the implementation overhead of the autoregressive verification layer.
What this means for multi-model architecture
The Hugging Face agent terminology publication, dated 25 May 2026, provides a useful operational frame: an agent is a model combined with a harness. The harness is the execution layer — model calls, tool handling, stopping conditions. The scaffold is the behavioural layer — system prompts, tool descriptions, context management. The direct implication: the same model in two different harnesses produces two distinct agent behaviours, per that publication.
This distinction becomes decisive in a multi-model architecture. If the harness is properly abstracted from the model provider, a specialised model can substitute a frontier model on a defined task without modifying the downstream pipeline. Conversely, if the harness is tightly coupled to a single vendor, every model decision carries a hidden migration cost that per-token price comparisons do not capture.
A coherent multi-model architecture rests on three layers: a specialised model on high-volume, well-defined tasks; a frontier model on exceptions and multi-domain tasks; an optimised inference engine on latency-constrained components. The harness layer is what makes this segmentation operable without a full rebuild at each vendor change.
Three levers to activate this week
- Identify a high-volume sub-domain in your current pipeline. If a frontier model is processing more than 100,000 homogeneous requests per month on a definable domain — extraction, classification, OCR — calculate the current cost and the projected cost with a 3B-to-7B specialised model. The 52× gap documented by Dharma-AI is an order of magnitude for calibrating the business case.
- Map your throughput bottlenecks. If your pipeline has latency or throughput constraints, test Nemotron-Labs diffusion mode on a real workload sample. The 6.4× gain published by NVIDIA is specific to self-speculation mode on B200 hardware — verify applicability to your infrastructure before any commitment.
- Audit your harness portability. Before any model decision, verify that your execution layer is abstracted from the model provider. If it is not, the true cost of each model arbitrage includes a migration cost that is invisible in the pricing comparison.
Is model size still the first criterion on your evaluation grid?
If this analysis speaks to you, I publish a piece of this calibre every day on digital innovation and enterprise AI. 👉 Get the next one straight in your inbox — sign-up takes ten seconds, and each edition is read before 9 a.m. by leaders of European SMEs, mid-caps and public institutions.