We're open-sourcing PulseBench-Tab, a frontier benchmark for table extraction. Table parsing remains one of the hardest and most poorly measured problems in document intelligence. TEDS operates on DOM trees and conflates HTML formatting conventions with structural errors. Needleman-Wunsch linearizes a two-dimensional structure into a one-dimensional sequence, so column transpositions can still score well because values align with nearby cells. GriTS uses greedy grid matching rather than optimal assignment and does not distinguish edge directions. The upshot: existing metrics cannot reliably separate content errors from structural errors, which makes provider comparisons noisy and downstream reliability unknowable. Alongside the dataset, our research team developed T-LAG. It parses each table into a cell-position grid, emits directed RIGHT and BELOW adjacency edges (suppressed within spanning cells, deduplicated by source, target, and direction), weights each candidate edge pair by the product of Levenshtein-derived similarities on source and target text, and uses the Hungarian algorithm for globally optimal one-to-one assignment. The F1 over matched edge weight is the T-LAG score. Structure and content are evaluated in one unified pass. HTML formatting choices do not affect the result. Rankings are invariant to the similarity exponent across k ∈ {7, 8, 9, 11}. The dataset contains 1,820 human-annotated tables across 9 languages and 4 scripts (Latin, CJK, Arabic, Cyrillic), drawn from 380 real-world financial filings, government reports, and regulatory disclosures. Tables range from 2 to 1,183 cells; 48.1% contain merged or spanning cells. Ground truth was produced through 8 annotation rounds with native speakers per language, independent cross-lingual review, and adversarial cell-by-cell audits against source images. We evaluated 9 commercial and open-source systems independently across the full dataset under exclude-missing scoring. Selected findings: @Pulse__AI Ultra 2 scores 0.9347 T-LAG; the next closest system scores 0.8155. Pulse Ultra 2 is the only provider with a median of 1.0, corresponding to perfect extraction on 57.9% of samples. Non-Latin scripts produce the widest cross-provider variance. On Arabic, the spread between top and bottom systems exceeds 75 percentage points. Structural hallucinations are pervasive. The second-ranked system achieves a perfect-extraction rate of 28.6%, meaning structural or content errors on 71.4% of tables (fabricated rows, invented content, incorrect span attributes, shifted data). Coverage failure is underreported. Multiple evaluated systems return no output on 19% to 21% of samples. Raw accuracy numbers without coverage disclosure favor selection bias. Thank you to Dushyanth Sekhar and Mohammed Hadi of S&P Global's Enterprise Data Organization for their academic contributions to the benchmark methodology. Dataset: huggingface.co/datasets/puls… Evaluation: github.com/Pulse-Software-Co… Blog: runpulse.com/blog/pulsebench… Research methodology: benchmark.runpulse.com/resea… Viewer: benchmark.runpulse.com

Another day, another benchmark broken. @Pulse__AI schema v2 scores #1 on LongExtractBench-50, the public set of the long-document extraction benchmark Reducto commissioned and Micro1 ran. Pulse scores 99.7 precision and 99.7 recall, run through Micro1's published harness and grader on the 50 released documents. This was done with 0 custom configuration on Pulse. Reducto, who commissioned the benchmark, published 99.3. One caveat: the leaderboard was scored on 225 documents and 50 are public. @micro1_ai , happy to run our models through your private set, or have your team run it as well!

Our schema extraction engine at @Pulse__AI is #1 on OmniExtractBench. We've rebuilt schema extraction from the ground up this summer, focusing on an agentic harness (more to share on this soon!) Last week it took #1 on ExtractBench, this week it is #1 on OmniExtractBench, Datalab's public benchmark. We ran their published scorer on the same 620 document set with 0 configuration settings on Pulse. Pulse scores 93.88, ahead of every system on the board (including the one that built it) - this is live on our public API now, try it!

Astra and Fable are genuinely excellent at reasoning over text, and dont get me started on how much ive enjoyed coding with these new models. the problem is that enterprise documents are not only text. The failures are the same on the models we test. Past ~50 pages, recall drops hard and the model does not notice, so it returns a confident result with most of the values missing. Schema logic, like "use the current period column" or "sum these rows unless a subtotal exists" starts drifting super quickly. The economics also dont help - running these complex/long docs through a coding agent costs 15 to 30 cents a page before retries, with no auditability on the output Each generation of oai/ant models is a large jump on reasoning and coding benchmarks , but a considerably smaller step on enterprise documents. Progress here comes from models trained on the problem directly

Awesome to see the latest results from our Schema V2 model. The price/accuracy curve has been a huge frontier push for our team, and Pulse non-effort is a sleeper: • Pulse Non-effort: 92.7 at $0.015/page + parsing Meanwhile: • LlamaExtract Agentic: 89.5 at ~$0.03/page • LlamaExtract Agentic Plus: 95.6 at ~$0.08/page • Reducto Deep Extract: 90.4 at ~$0.05/page And if you want maximum accuracy, Pulse Effort reaches 95.9 at $0.09/page + parsing. Try it out now!

A few months ago, the team at @Pulse__AI launched PulseBench-Tab, our frontier table extraction benchmark with 1,820 human-annotated tables across 9 languages. Our evals team has updated the benchmark against the latest generation of document parsers and frontier models. Pulse Ultra 2 remains state of the art on the benchmark, and is now just 1 credit per page. As always, everything is open source: the scorer, the runners for every provider, the dataset, and every prediction. Leaderboard: benchmark.runpulse.com GitHub: github.com/Pulse-Software-Co… Dataset: huggingface.co/datasets/puls… PRs welcome if you'd like a provider added or a sample corrected. Excited to keep pushing the frontier of table extraction.

. @reductoai released their latest model this week, r-1. The @pulse__ai team ran it on PulseBench-Tab, our frontier table-extraction benchmark consisting of thousands of human-annotated, multilingual tables across 9 languages. Pulse Ultra 2 remains state-of-the-art with r-1 taking minimal accuracy improvements over their previous generation. Pulse Ultra 2: 0.935 Reducto r-1: 0.811 Reducto Agentic: 0.795 We will be rolling out PulseBench-Tab updates throughout the week, DM us if you’d like a parser included!

I'm super excited to share that we have raised $250M at a $2.3B valuation to build a LOT more AI compute capacity in orbit 🚀 Thanks so much to our supporters, including Manhattan West, who led the round, and new investors, @NVIDIA, @Cisco_Invests, @CedarCapital, Goanna Capital, and @Standard_Cap, as well as continued investment from existing backers @Benchmark, @EQTVentures, @Soma_Capital, @NFX, @SevenSevenSix, and others. Next week, @Starcloud_ will move into a new 100,000 square ft facility that will enable us to ramp up our production rate to 100 satellites/week. This is just the beginning! @EzraFeilden, @AdiOltean🏃🏻‍➡️🏃🏻‍➡️🏃🏻‍➡️

Our engineering team has been launching a lot of features. Here's what we launched over the past few weeks at @Pulse__AI 1/ MCP server 2/ Generate-split 3/ Selection model 4/ CLI 5/ One-click SSO 6/ Tutorial mode All of it is live today, and there's more on the way.

Charts are among the most valuable objects in an enterprise document, and until now they have also been among the least usable. A decade of company performance can live inside a single line chart, and decades of subsurface measurement inside a scanned well log. Conventional OCR returns the title, the legend, and the surrounding text, because a chart stores most of its meaning in the height of a bar or the slope of a line rather than in any character on the page. Today we are releasing @Pulse__AI Agentic Chart Reconstruction to the general public. It combines a family of small, specialized vision models with a VLM that orchestrates them, deciding where the system needs to look more closely. Clean digital charts move through the fast path, while faded scans, overlapping series, and tick labels too small for a full-page pass receive additional attention before anything is returned. The output is structured chart data with the context required to use it, including series names, colors, line styles, axis values, legends, and the location of the chart in the source document, so every recovered number can be checked against the figure it came from. For historical well logs, a single degraded scan can yield thousands of values aligned to a shared depth grid, exported as JSON, Excel, CSV, or LAS files that move directly into the software subsurface teams already run. The system has been in beta with a group of leading oil and gas companies recovering data from historical logs and technical records, and it is now available through the Pulse platform and API for line, scatter, bar, pie, donut, and well-log charts. The earlier generation of document systems made text searchable; Pulse makes the entire document usable. Try it now with your hardest charts! runpulse.com/blog/pulse-rele…

Pulse turns complex documents such as financial statements, contracts, and claims into clean markdown and structured data for agents. With MPP, your agent can use Pulse as soon as a task requires it; payment and access to the data are handled through the API and settled on Tempo

Your agent needs data it can actually work with. Pulse handles that; turning complex documents like financial statements, contracts, and claims into clean markdown and structured data for agents. With MPP, agents can access and use Pulse via API. No accounts needed:

Agentic payments are now live on @Pulse__AI . An agent can point at our API with no account, no API key, and no signup, and pay for extraction on its own, one call at a time. It runs on the Machine Payments Protocol (MPP), settling inline over HTTP in USDC, in the same request that asks for the work. The agent sends an extraction request, Pulse answers 402 Payment Required with the priced terms, the agent pays and retries with proof, and the result comes back. Two round trips, no invoice to reconcile. The 402 has sat reserved since the early days of HTTP, waiting for a caller that could pay on its own. It’s live on Pulse today, more here: runpulse.com/blog/welcome-to…

@Pulse__AI ran GDP.pdf through their document stack. GDP.pdf measures whether AI can handle the messy, high-stakes documents enterprises run on. Pulse kept the models and the grader fixed and changed only the evidence layer - feeding each model a cleaner, structured version of the PDF. The result was +10 points in whole-task accuracy. One of the things they noticed: models often fail because a footnote gets dropped, a value comes from the wrong row, or a number loses its unit along the way. The public set is open to anyone working on document AI. Check it out! huggingface.co/datasets/surg…

Every document extraction system can read the text on a form, yet almost none of them can read the pen, so the checkmarks, circles, underlines, and strikethroughs people use to record decisions get treated as noise and quietly ignored. Today Pulse is launching Selection Mark, a model trained to read them, available now through a single parameter in the extraction API. More: runpulse.com/blog/capturing-…

Surge's GDP.pdf benchmark tests the documents enterprises run on, and even top frontier models pass only a fraction of whole tasks. The bottleneck isn't model intelligence. It's input quality. @Pulse__AI as the extraction layer lifted whole-task success from ~22% to ~32% (criteria-level ~80%) on the same models. When retrieval preserves document structure, models reason over the right facts. Full write-up here: runpulse.com/blog/gdp-pdf-an…

Every enterprise that processes documents eventually hits the same wall. Extracting data from PDFs is one thing, but putting data back into a PDF is where the workflow breaks down. Most enterprise PDFs don't have defined form fields at all. Government forms arrive as scanned images, insurance claims are flat PDFs, and healthcare intake forms have been printed and photocopied beyond recognition. Traditional tools have nothing to work with. @Pulse__AI Form Fill uses our vision-language models to analyze the full document layout, detect where fillable regions exist based on visual context, and place values in the correct locations. You provide structured data or natural language instructions, and Pulse returns a filled PDF. Healthcare teams are already using this to automate pre-auth forms, CMS-1500 claims, and provider enrollment packets through a single API call. Full blog post: runpulse.com/blog/introducin…