AI Models for Research and Documents: Gemini, GPT and Claude versus Open LLMs

A readable summary can omit important details or cite the wrong sources. Every step matters: reading the file, recognizing text and images, finding relevant passages and checking claims against the original. This comparison describes documented features of Gemini 3.8 Flash, GPT-6.1 Sol, Claude Opus/Sonnet 5.5, DeepSeek V4.1 Flash and Qwen3.8-27B. It does not report an original quality test of these models. Sources were checked on 25 October 2026.

Gemini and document ingestion

Gemini 3.8 Flash documents support for PDF, text, image, video and audio inputs. File search and web search are additional supported tools. Search grounding means basing answers on retrieved sources. The application needs to provide and use those tools. For PDFs, Gemini can also process visible tables and illustrations; file size, page count and input limits still apply. Support for a file format does not establish that every footnote was read correctly. [3, 7]

GPT and Claude under the same protocol

GPT-6.1 Sol, Claude Opus 5.5 and Sonnet 5.5 support text and image inputs. [1, 2] A fair comparison requires the same documents and questions. The test record also needs to identify the search tools used. An answer following web search cannot be directly compared with one based only on model knowledge. Separate tests can reveal whether an error arose during source retrieval or while reading and summarizing.

Additional components for open models

DeepSeek V4.1 Flash and Qwen3.8-27B publish model weights for text and image processing. Weights are the stored parameters of a trained model. Downloading weights does not provide a complete document search system. A parser reads file structure; OCR converts scanned writing into text. A search index makes a collection searchable, and retrieval selects relevant passages. Not every workflow requires all these steps. [4, 5] Fully local processing requires these supporting services to run locally too. Ollama also offers cloud models and web search, so the program name alone does not establish a local data path. The model and runtime must both support the required image processing. [6]

A test with verifiable answers

A proposed test package contains a short report, a long PDF with tables, a scan and two versions containing conflicting information. One task could ask for revenue from a particular table row, together with its page and row location. Other tasks assess summaries, contradictions and a question whose answer is absent from the documents. Expected facts and source locations are recorded before testing. This package is a proposal, not a completed experiment reported here.

The record counts correct statements, missing required facts and invented evidence separately. An unavailable answer must remain identifiable as unavailable. Checking a quoted number also requires its unit, period and table heading. A link to the correct website is insufficient when the figure appears only in another edition or is absent altogether. Repeated runs can also reveal whether results remain consistent.

The editorial shortlist

Gemini is a candidate for mixed files because of its documented input formats. GPT and Claude are further candidates for the same document test. Models with published weights allow a self-hosted workflow when hardware, software and usage terms permit. That establishes neither better quality nor automatically confidential processing. A sound choice considers correct answers, omissions, runtime, cost and the actual data path.

Sources

1. OpenAI: GPT-6.1 Sol
2. Anthropic: Claude model overview
3. Google: Gemini 3.8 Flash
4. DeepSeek: V4.1 Flash model card
5. Qwen: Qwen3.8-27B model card
6. Ollama: FAQ and local/cloud operation
7. Google: document understanding

Sources checked: 25 October 2026.