{"id":4279,"date":"2026-09-25T17:59:46","date_gmt":"2026-09-25T08:59:46","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/model-tencent-wevisdoc-en\/"},"modified":"2026-09-28T05:32:20","modified_gmt":"2026-09-27T20:32:20","slug":"model-tencent-wevisdoc-en","status":"publish","type":"page","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-tencent-wevisdoc-en\/","title":{"rendered":"WeVisDoc Guide: VRAM Requirements, GGUF Builds"},"content":{"rendered":"<h2>About This Model<\/h2>\n<p><strong>WeVisDoc<\/strong> is a model from Tencent that <strong>turns images of document pages into structured text.<\/strong> Give it a PDF page or a scanned image and it returns Markdown that preserves headings and paragraph structure, with formulas as LaTeX and tables as HTML. Rather than chaining separate tools for layout detection, text recognition and table recognition, it does the whole job <strong>end-to-end in a single model.<\/strong><\/p>\n<p>It is built by adapting Qwen3-VL (2B and 4B), a language model that can read images, for document parsing, and comes in two sizes: <strong>WeVisDoc-2B<\/strong> and <strong>WeVisDoc-4B<\/strong>.<\/p>\n<h2>What Makes It Stand Out<\/h2>\n<p>The publisher reports that it ranks first among the compared end-to-end document parsers in all four evaluation settings. An excerpt:<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Model<\/th>\n<th style=\"text-align: right;\">Size<\/th>\n<th style=\"text-align: right;\">OmniDocBench v1.6 overall \u2191<\/th>\n<th style=\"text-align: right;\">PureDocBench real-world degraded \u2191<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>WeVisDoc-4B<\/strong><\/td>\n<td style=\"text-align: right;\">4B<\/td>\n<td style=\"text-align: right;\"><strong>95.38<\/strong><\/td>\n<td style=\"text-align: right;\"><strong>69.08<\/strong><\/td>\n<\/tr>\n<tr>\n<td><strong>WeVisDoc-2B<\/strong><\/td>\n<td style=\"text-align: right;\">2B<\/td>\n<td style=\"text-align: right;\">95.06<\/td>\n<td style=\"text-align: right;\">65.60<\/td>\n<\/tr>\n<tr>\n<td>Logics-Parsing-v2<\/td>\n<td style=\"text-align: right;\">4B<\/td>\n<td style=\"text-align: right;\">93.33<\/td>\n<td style=\"text-align: right;\">67.64<\/td>\n<\/tr>\n<tr>\n<td>FD-RL<\/td>\n<td style=\"text-align: right;\">4B<\/td>\n<td style=\"text-align: right;\">91.21<\/td>\n<td style=\"text-align: right;\">67.04<\/td>\n<\/tr>\n<tr>\n<td>HunyuanOCR-1.5<\/td>\n<td style=\"text-align: right;\">1B<\/td>\n<td style=\"text-align: right;\">94.74<\/td>\n<td style=\"text-align: right;\">61.59<\/td>\n<\/tr>\n<tr>\n<td>Qianfan-OCR<\/td>\n<td style=\"text-align: right;\">4B<\/td>\n<td style=\"text-align: right;\">93.90<\/td>\n<td style=\"text-align: right;\">45.06<\/td>\n<\/tr>\n<tr>\n<td>dots.ocr<\/td>\n<td style=\"text-align: right;\">3B<\/td>\n<td style=\"text-align: right;\">90.77<\/td>\n<td style=\"text-align: right;\">55.68<\/td>\n<\/tr>\n<tr>\n<td>DeepSeek-OCR 2<\/td>\n<td style=\"text-align: right;\">3B<\/td>\n<td style=\"text-align: right;\">90.25<\/td>\n<td style=\"text-align: right;\">43.60<\/td>\n<\/tr>\n<tr>\n<td>olmOCR-2-7B<\/td>\n<td style=\"text-align: right;\">7B<\/td>\n<td style=\"text-align: right;\">85.51<\/td>\n<td style=\"text-align: right;\">56.10<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<ul>\n<li><strong>On clean documents, the top models are now close.<\/strong> The leaders cluster at 94\u201395 overall on OmniDocBench, and WeVisDoc&#8217;s lead is under one point.<\/li>\n<li><strong>Models separate on photographed and degraded documents.<\/strong> On PureDocBench&#8217;s &#8220;Real Degraded&#8221; track, Qianfan-OCR\u2014close on clean documents\u2014drops to 45.06 while WeVisDoc-4B holds 69.08. Some models do come close here, though: Logics-Parsing-v2 (67.64) and FD-RL (67.04) are within about 1.4\u20132 points. The accurate reading is &#8220;the best among the models that handle degraded documents well.&#8221;<\/li>\n<li><strong>On clean documents, the 2B version nearly matches the 4B.<\/strong> Its overall 95.06 is close to the 4B&#8217;s 95.38. On degraded documents, however, it scores 65.60, behind the 4B and the two models above. If you mostly process phone photos or old scans, the 4B version is the safer choice.<\/li>\n<\/ul>\n<p>Some baseline scores were re-measured by the publisher; the rest are taken from each model&#8217;s paper.<\/p>\n<h2>Running It Locally<\/h2>\n<ul>\n<li><strong>At 4B and 2B, it is small enough for common GPUs.<\/strong> The official instructions run it as a vLLM server that you send page images to. This site also tracks converted builds such as GGUF (see &#8220;Quantized and Converted Variants&#8221; below).<\/li>\n<li>It is specialized for reading documents; it is not a general-purpose chat model.<\/li>\n<li><strong>The license is Apache-2.0, which allows commercial use.<\/strong> The technical report is arXiv:2609.20423.<\/li>\n<\/ul>\n<p><em>Source: <a href=\"https:\/\/huggingface.co\/tencent\/WeVisDoc-4B\">tencent\/WeVisDoc-4B model card<\/a>, as of 2026-09-25. Benchmark figures are as published by Tencent.<\/em><\/p>\n<h2>Our Coverage and Data<\/h2>\n<p>Everything Local Model Watch has published about the <strong>WeVisDoc<\/strong> family: 1 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-en\/\">model family index<\/a>.<\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Base model(s)<\/td>\n<td><a href=\"https:\/\/huggingface.co\/tencent\/WeVisDoc-4B\">tencent\/WeVisDoc-4B<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-tencent-en\/\">Tencent<\/a><\/td>\n<\/tr>\n<tr>\n<td>Parameters<\/td>\n<td>4.4B<\/td>\n<\/tr>\n<tr>\n<td>License (model card)<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Smallest VRAM tier<\/td>\n<td>4GB<\/td>\n<\/tr>\n<tr>\n<td>Articles<\/td>\n<td>1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong> \u2014 4.4B parameters<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>4GB (laptop iGPU \/ phone class)<\/td>\n<td>Q5_K_S<\/td>\n<td>2.9GB<\/td>\n<td>3.5GB<\/td>\n<\/tr>\n<tr>\n<td>8GB (RTX 4060 \/ 3060 Ti, etc.)<\/td>\n<td>Q8_0<\/td>\n<td>4.4GB<\/td>\n<td>5.2GB<\/td>\n<\/tr>\n<tr>\n<td>12GB (RTX 4070 \/ 3060 12GB, etc.)<\/td>\n<td>F16<\/td>\n<td>8.2GB<\/td>\n<td>9.9GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. File sizes are measured from the converted build <a href=\"https:\/\/huggingface.co\/mradermacher\/WeVisDoc-4B-GGUF\">mradermacher\/WeVisDoc-4B-GGUF<\/a>. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p>The publisher distributes this model as safetensors.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><strong>Compression:<\/strong> the Q5_K_S build measures 5.61 bits per weight \u2014 about 35% the size of the original 16-bit weights, calculated by this site from the actual file sizes.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats and the license field. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<h2>Quantized and Converted Variants<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Added<\/th>\n<th>Publisher<\/th>\n<th>Format<\/th>\n<th>Repository<\/th>\n<th>Smallest VRAM tier (build, est. memory)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>2026-09-20<\/td>\n<td>mradermacher<\/td>\n<td>GGUF<\/td>\n<td><a href=\"https:\/\/huggingface.co\/mradermacher\/WeVisDoc-4B-GGUF\">mradermacher\/WeVisDoc-4B-GGUF<\/a><\/td>\n<td>Q5_K_S 3.5GB (fits in 4GB VRAM)<\/td>\n<\/tr>\n<tr>\n<td>2026-09-20<\/td>\n<td>mradermacher<\/td>\n<td>GGUF (imatrix)<\/td>\n<td><a href=\"https:\/\/huggingface.co\/mradermacher\/WeVisDoc-4B-i1-GGUF\">mradermacher\/WeVisDoc-4B-i1-GGUF<\/a><\/td>\n<td>Q5_K_S 3.5GB (fits in 4GB VRAM)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>File sizes of each build:<\/p>\n<ul>\n<li>Available builds in mradermacher\/WeVisDoc-4B-GGUF: Q2_K 1.7GB \/ Q3_K_S 1.9GB \/ Q3_K_M 2.1GB \/ Q3_K_L 2.2GB \/ IQ4_XS 2.3GB \/ Q4_K_S 2.4GB \/ Q4_K_M 2.5GB \/ Q5_K_S 2.9GB \/ Q5_K_M 2.9GB \/ Q6_K 3.4GB \/ Q8_0 4.4GB \/ F16 8.2GB<\/li>\n<li>Available builds in mradermacher\/WeVisDoc-4B-i1-GGUF: IQ1_S 1.1GB \/ IQ1_M 1.2GB \/ IQ2_XXS 1.3GB \/ IQ2_XS 1.4GB \/ IQ2_S 1.5GB \/ IQ2_M 1.6GB \/ Q2_K_S 1.6GB \/ Q2_K 1.7GB \/ IQ3_XXS 1.7GB \/ IQ3_XS 1.8GB \/ Q3_K_S 1.9GB \/ IQ3_S 1.9GB \/ IQ3_M 2.0GB \/ Q3_K_M 2.1GB \/ Q3_K_L 2.2GB \/ IQ4_XS 2.3GB \/ Q4_0 2.4GB \/ IQ4_NL 2.4GB \/ Q4_K_S 2.4GB \/ Q4_K_M 2.5GB \/ Q4_1 2.6GB \/ Q5_K_S 2.9GB \/ Q5_K_M 2.9GB \/ Q6_K 3.4GB<\/li>\n<\/ul>\n<p><em>This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.<\/em><\/p>\n<h2>Articles (the family&#8217;s own models first, then newest)<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Published<\/th>\n<th>Model<\/th>\n<th>Type<\/th>\n<th>Article<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>2026-09-18<\/td>\n<td>tencent\/WeVisDoc-4B<\/td>\n<td>New Models<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/18\/tencent-releases-wevisdoc-document-parsing-models\/\">Tencent Releases WeVisDoc-2B and WeVisDoc-4B Document Parsing Models<\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<h2>Repositories<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/tencent\/WeVisDoc-4B\">tencent\/WeVisDoc-4B<\/a><\/li>\n<\/ul>\n<p><em>Last updated 2026-09-28 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under &#8220;Our Coverage and Data&#8221; are assembled by code from our article log.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>About This Model WeVisDoc is a model from Tencent that turns images of document pages into structured text. Give it a PDF page or a scanned image and it returns Markdown that preserves headings and paragraph structure, with formulas as LaTeX and tables as HTML. Rather than chaining separate tools for layout detection, text recognition [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":""},"class_list":["post-4279","page","type-page","status-publish","hentry"],"lang":"en","translations":{"en":4279,"ja":4278},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages\/4279","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=4279"}],"version-history":[{"count":5,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages\/4279\/revisions"}],"predecessor-version":[{"id":5990,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages\/4279\/revisions\/5990"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=4279"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}