Last Week in AI for Chemicals:
5 Oct to 11 Oct 2026

Read in depth.

New test results.

Public tests score AI models on set tasks, such as answering regulatory questions. These are last week's results that matter for chemical companies.

1. 9 Oct 2026

   LightOnOCR-3, a Free Document Reader, on ParseBench

   LightOnOCR-3 is a document reader: it turns page images into text and tables. LightOn released it on 8 October as three small models, free for commercial use under the Apache 2.0 license, so you can run it on your own computers. On ParseBench the largest, at 4 billion parameters, scores 75.1 out of 100 on whole documents, 16th of 147. LightOn submitted that result itself on 9 October, and the company that runs ParseBench sells a rival tool. OmniDocBench, the guide's independent document test, hasn't scored it, so TeleOCR stays the guide's pick for reading your TDSs and SDSs on your own computers.

   ParseBench leaderboard9 Oct 2026

   LightOn: LightOnOCR-38 Oct 2026

   Read Supplier Documents in the guide →
2. 6 Oct 2026

   Mistral Large 4 Lands Mid-Table on Regulatory Research

   Mistral Large 4 is the largest general model from Mistral, a French AI company, and it went out as a public preview on 6 October. It answers 34.85% of US regulatory research questions fully right, 47th of 76 on the test, at $5.12 a question; Claude Sonnet 5.5 leads that test at 69.7%. On real professional tasks it ranks 19th of the 25 models Artificial Analysis rates. Mistral says the weights, the files you need to run it on your own servers, follow by the end of October. Until then it runs only on Mistral's own service, at $1.36 for every million tokens sent and $4.18 for every million written, at half price for its first two weeks.

   vals.ai Legal Research, regulatory questions12 Oct 2026

   Artificial Analysis: models and the Intelligence IndexNo date stated

   Mistral AI: Mistral Large 46 Oct 2026

   Research Regulations in the guide →

Released, not yet tested.

No independent test has run these, so the guide shows no score.

• Reflection Beam, an Open Model Still to Come

  No independent test

  Beam is a large general AI model, the kind that writes, summarizes, and reasons through a question, from Reflection AI. Reflection announced it on 5 October and says it will release Beam free to download under the Apache 2.0 license later this month, so a company could run it on its own servers. For now access is by waitlist, there is no public price, and every score so far is Reflection's own, so the guide shows none.

  Published 5 Oct 2026

  Reflection AI: Introducing Beam5 Oct 2026

• Cloudflare Clef-omni, a Free Model That Sorts Requests

  No independent test

  Clef-omni is a model that doesn't write a reply. It reads an email, a form, an image, a recording, or a video and picks one answer from a list you set, such as which team an RFQ should go to. Cloudflare released it on 9 October, free to download under the Apache 2.0 license, and runs it on its own service at $0.15 for every million tokens sent. No public test has measured it.

  Published 9 Oct 2026

  Cloudflare: Introducing Clef-omni9 Oct 2026

• OpenAI Opens Its Decisions API

  No independent test

  OpenAI's Decisions API is a service that reads text or an image and returns one set answer, such as a category or a yes or no, instead of writing a reply. It opened as a public beta on 6 October, runs only on GPT-6 Luna, and costs $0.10 for every million tokens sent, with no charge for the answer. It suits jobs such as sorting your RFQs and order emails, and no public test has measured it.

  Published 6 Oct 2026

  OpenAI: the Decisions APINo date stated

• Liquid AI's d1 Decision Models, Free to Download

  No independent test

  d1-3B and d1-omni-600M are two small models from Liquid AI that read text or images and pick one answer from a list you set, such as what kind of document a file is. Liquid AI published them on 7 October, free to download under its own license rather than a standard open source one. No test the guide uses has scored them.

  Published 7 Oct 2026

  Liquid AI: Open d17 Oct 2026

• Tencent Youtu-Parsing-Omni, a Document Reader to Download

  No independent test

  Youtu-Parsing-Omni is a document reader: it turns a scanned page, a chart, a photo, or a recording into structured data a system can use. Tencent published it on 9 October. You can download it and run it on your own computers, under Tencent's own license rather than a standard open source one. Every score so far is Tencent's own, so the guide shows none.

  Published 9 Oct 2026

  Tencent: Youtu-Parsing-OmniNo date stated

• EmbeddingGemma 2, a Free Search Model From Google

  No independent test

  EmbeddingGemma 2 is a search model. It doesn't answer questions; it turns text, images, and other files into a form a search tool can match, so a search can find the right TDS even when the wording differs. Google released it on 6 October, free under the Apache 2.0 license, and it's small enough to run on one computer. Only Google's own scores exist so far.

  Published 6 Oct 2026

  Google: EmbeddingGemma 26 Oct 2026

Still no public test.

Jobs where nobody published a test last week. Knowing what has not been measured tells you what to check yourself.

• Read Supplier Documents

  No public test last week measured reading a CoA, a TDS, or an SDS with current models.

• Research Regulations

  No public test last week asked about GHS, CLP, REACH, TSCA, or dangerous goods rules. The guide's regulatory test is US administrative law.

• Sort Requests and RFQs

  No public test last week measured sorting RFQs or order emails, though three decision model releases are built for it.

• Screen Sanctions Lists

  No public test last week measured sanctions screening with current models.