Sissach, 28 September 2026<\/p>\n
Parashift today releases version 5.3 of Dokumentli\u00ae<\/a>, the specialised vision language model for processing European business documents. It replaces launch version 5.2 of 8 September 2026 and is available immediately for download and in the playground.<\/p>\n The benchmark report<\/a> of 25 September 2026 measures Dokumentli\u00ae 5.3 across 3,365 real business documents from five areas: invoice processing, e-commerce, logistics, automotive and real estate.<\/p>\n Accuracy rises from 88.4 to 89.1 percent<\/strong> on cANLS@0.8. Throughput rises from 0.86 to 1.14 documents per second<\/strong>, an increase of 33 percent.<\/p>\n Dokumentli\u00ae 5.3 wins all five benchmarks against Claude Opus 5.5, GPT-6 Astra, Gemini 3.8 Flash, Qwen3.8-27B and Gemma 4 26B. It does so with a context budget of 8,192 tokens, while the reference models use up to 64,000 tokens.<\/p>\n Datasets and reference models were adjusted compared with the August measurement.<\/p>\n Dokumentli\u00ae is developed continuously. New model versions are included in the subscription and roll out without additional cost and without migration effort.<\/p>\n Dokumentli\u00ae 5.3 runs unchanged on a single GPU with 24 GB of VRAM, on premises in your own data centre or in sovereign processing zones in Zurich, Frankfurt and Amsterdam. Customer data is never used for training.<\/p>\n Extraction is controlled through a JSON specification, without templates and without form-specific training.<\/p>\n Try Parashift on your own documents: test it in the playground<\/a>, book a demo<\/a> or get in touch<\/a>.<\/p>\n","protected":false},"excerpt":{"rendered":" Version 5.3 of Dokumentli\u00ae replaces launch version 5.2. In the benchmark of 25 September 2026 the model reaches 89.1 percent accuracy and 1.14 documents per second.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"content-type":"","footnotes":""},"categories":[105],"tags":[],"class_list":["post-61251","post","type-post","status-publish","format-standard","hentry","category-press-release"],"_links":{"self":[{"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/posts\/61251","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/comments?post=61251"}],"version-history":[{"count":1,"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/posts\/61251\/revisions"}],"predecessor-version":[{"id":61252,"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/posts\/61251\/revisions\/61252"}],"wp:attachment":[{"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/media?parent=61251"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/categories?post=61251"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/parashift.ai\/en\/wp-json\/wp\/v2\/tags?post=61251"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Results from the September benchmark<\/h2>\n
Continuous model improvements<\/h2>\n
Deployment and availability<\/h2>\n
Related<\/h3>\n
\n