ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #OCRia-1.5 #LlamaParse #Document Parsing #Visual Grounding #ParseBench-2

LlamaIndex Launches OCRia-1.5: 2B and 4B Models for More Accurate Document Parsing

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:LlamaIndex has launched OCRia-1.5 in LlamaParse. The upgraded 4B model powers the Agentic tier, while a new 2B model serves the Cost Effective tier, with improvements focused on complex tables, charts, forms, and bounding-box localization. On LlamaIndex's ParseBench-2, the Agentic table score rises from 86.5 to 90.1, the Cost Effective chart score from 67.3 to 73.0, and the Cost Effective visual-grounding score from 77.0 to 88.5. This is a managed-service update: architecture details, weight av


LlamaIndex has put OCRia-1.5 into production across LlamaParse. The release uses one model generation to cover two service paths: the upgraded 4B model powers Agentic, while the new 2B model serves Cost Effective, with complex tables, charts, forms, and bounding-box localization as the main targets.

Release Scope: Two Sizes for Different Accuracy-Cost Paths

  • OCRia-1.5-4B is the upgraded flagship for the Agentic tier, aimed at complex tables, charts, and higher-accuracy workloads.
  • OCRia-1.5-2B has roughly half the parameter count of the previous 4B OCRia-1 model and serves the Cost Effective tier for high-volume parsing.
  • Both models are live in LlamaParse. The announcement does not provide open weights, a model license, or a self-hosting route, so this is a managed-product release rather than an open-source launch.

Agentic appears to be a LlamaParse service tier rather than proof of a specific model architecture. The post does not establish that the 4B model itself uses autonomous planning, multi-tool calls, or iterative reasoning; the tier may also include an orchestration layer. Model capability and end-to-end service capability therefore need to be evaluated separately.

Technical Mechanism: Preserving Structure Rather Than Merely Reading Characters

Complex document parsing is not just OCR. Extracted characters must be placed back into the correct structural hierarchy. A heading such as Three Months Ended may span several columns in a financial statement. If the visual layout is flattened, the numbers can remain correct while their reporting period is silently misassigned.

OCRia-1.5 targets three principal output classes. The first is table structure, including cells, multi-row or multi-column headers, and their relationships. The second is chart semantics, requiring axes, legends, series, and labels to remain associated. The third is form state, which requires preserving not only field values but also checkbox status. LlamaIndex says gains were largest for tables with spanning headers, although it does not disclose the size or scores of that subset.

The fourth target is visual grounding. Bounding boxes bind extracted results back to source regions, allowing RAG and agent workflows to highlight supporting evidence rather than emit fields detached from the page. Boxing each form field and covering each complete table can reduce neighboring-text mix-ups and missing evidence. However, the announcement does not reveal the visual encoder, input resolution, token budget, training losses, synthetic-data strategy, or distillation pipeline. It is therefore impossible to attribute the gains specifically to model scale, higher visual resolution, a new layout objective, or service orchestration.

Engineering Trade-offs: A Smaller Model Does Not Guarantee Twice the Throughput

If the variants share an architecture and numerical precision, the 2B model should require roughly half the weight storage of the 4B model. That does not imply a twofold improvement in end-to-end parsing speed. Page rendering, image encoding, visual tokenization, OCR decoding, structural post-processing, and network transfer can account for substantial latency. The announcement provides no p50 or p95 latency, concurrency, throughput, or memory-footprint data.

The 4B model is stronger on complex tables and charts, but aggregate scores do not prove that it is better on every document class. A practical production design routes by complexity: ordinary pages use the 2B model, while pages with spanning headers, dense charts, or irregular forms are promoted to the 4B model instead of applying flagship cost to the entire corpus.

The Cost Effective tier is listed at 3 credits per page, while the 4B model's Agentic tier is listed at 10 credits per page. The original lead also states that Agentic Plus costs 25 credits per page, but the body does not clearly map that tier to the two models. Buyers should confirm the live billing configuration in the console.

A more useful cost model is:

Total page cost = base parsing credits + retry or Agentic surcharges + any other OCR and post-processing cost

Teams should also compare cost per successfully structured field or per verifiable evidence item rather than relying on price per page alone.

Vendor-Reported Results: ParseBench-2 Score Changes

| Metric | Service tier | Previous | New | Delta | |---|---:|---:|---:|---:| | Tables | Agentic | 86.5 | 90.1 | +3.6 | | Tables | Cost Effective | 83.2 | 87.5 | +4.3 | | Charts | Agentic | 74.9 | 76.8 | +1.9 | | Charts | Cost Effective | 67.3 | 73.0 | +5.7 | | Visual grounding | Cost Effective | 77.0 | 88.5 | +11.5 | | Visual grounding | Agentic | Not disclosed | Not disclosed | +4.1 | | Forms | Agentic | Not disclosed | Not disclosed | +2.6 | | Forms | Cost Effective | Not disclosed | Not disclosed | +4.2 |

The larger improvements in Cost Effective chart and visual-grounding scores show that the smaller model does not merely trade accuracy for lower cost. LlamaIndex also says the Cost Effective tier is now more accurate on all six ParseBench-2 dimensions, but it does not publish a complete table of absolute scores for all six.

Benchmark Credibility: Vendor Results Still Require Independent Analysis

ParseBench-2 is maintained by LlamaIndex, so its claims about state-of-the-art table, chart, and formatting performance should be treated as first-party results. The announcement does not disclose sample size, data splits, scoring formulas, confidence intervals, run counts, configuration differences, or version-locked baselines.

The benchmark evaluates LlamaParse tiers rather than isolated models, so score changes can include effects from the model, prompts, preprocessing, post-processing, and service orchestration. The visual-grounding score also cannot automatically be interpreted as pixel-level IoU: the post does not specify thresholds, empty-box handling, cross-region boxes, or normalization across page sizes.

Any leadership claim needs third-party reproduction under matched page rendering, OCR post-processing, concurrency, and cost conditions. Enterprise evaluations should also report results by scan quality, language, table complexity, merged cells, and form type, because a single aggregate score can conceal long-tail failures.

Developer Deployment: Route Documents by Complexity

  1. Build a stratified test set: Include financial statements, merged-cell tables, multi-chart pages, invoices, handwritten annotations, scanned documents, and multilingual files instead of relying on random office paperwork.
  2. Separate quality metrics: Independently evaluate header-to-cell relationships, chart-series labels, form-field and checkbox state, bounding-box coverage, and localization offset. One aggregate score can hide structural errors.
  3. Measure full service cost: Track credits per page, retry rates, Agentic surcharges, p50 and p95 latency, concurrent throughput, and fallback costs.
  4. Implement two-tier routing: Use the 2B model as the high-volume default and promote complex tables and charts to the 4B model. Escalate or send to human review when structural confidence remains low rather than writing questionable output directly into a knowledge base.
  5. Validate downstream citations: For RAG and agent systems, check whether citations land on the correct field or table region. More accurate boxes do not prevent a generative model from misreading the extracted content.
  6. Review the data boundary: Cloud parsing uploads source documents, so teams should confirm data residency, log retention, zero-retention options, subprocessors, and DPA terms. Since weights and local deployment are not provided, this release alone does not satisfy strict on-premises requirements.

OCRia-1.5's practical value lies in using 2B and 4B service paths to balance high-volume and complex-document workloads while reconnecting structured output to page-level evidence. Its main limitations remain limited architectural transparency, missing latency data, and the absence of independent evaluation.


Source: LlamaIndex Blog (2026-11-18)

— END —

Tags: #OCRia-1.5 #LlamaParse #Document Parsing #Visual Grounding #ParseBench-2

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…