Open Models This Month: DeepSeek’s Update, a 1 GB Assistant Model, and a Reader for 100-Page Documents

DeepSeek-V4.1-Flash-1017 ranks first among 23 open-weight models of similar size on Artificial Analysis, with a score of 32. Liquid AI’s 1.7-billion-parameter model fits in about 1 GB, and Baidu’s reader handles 100-page documents in one pass.
artificial-intelligence
data-science
Author

Kabui, Charles

Published

2026-11-02

Keywords

open-weights, deepseek, on-device-ai, document-ocr, speech-to-text, agent-benchmarks

Download as PDF

DeepSeek-V4.1-Flash-1017, released October 17, is a downloadable open-weight model for writing software and handling multi-step tasks, with 552 billion parameters (the numbers a model learns, roughly its size). DeepSeek reports 42.3 on Terminal-Bench 4.0 (command-line tasks), up from 31.2 before the update. The independent tracker Artificial Analysis combines nine tests into one index and ranks it first among 23 similar-size open-weight models, at 32 against 59 for Gemini 4 Carbon. GLM-5.3 ties it at 32 but costs $5.01 per test task, against $0.60 for DeepSeek.

Smaller open-weight releases fit one device or job. Liquid AI’s LFM3-1.7B is a text model for assistants that run on your own device. It needs about 1 GB of memory when compressed to 4-bit numbers, and Liquid says it beats its 2.6-billion-parameter predecessor. Baidu’s Unlimited-OCR-2 turns scanned pages into text, reading a 100-page document in one pass. Baidu reports it needs 1.77 GB of short-term memory, against 47 GB with the usual method. NVIDIA’s Nemotron 3.5 Transcribe turns hour-long recordings into text in up to 40 languages, but not live speech.

Baidu’s test for long-document readers, MarathonDocBench, holds 100 documents of 100 pages or more. Unlimited-OCR-2 averages 77.06 out of 100 across eight accuracy measures and the next model listed scores 33.7, but Baidu wrote the test and lists no independent results. BizAgentBench gives AI agents 50 payment-company tasks and 601 business tools. The September DeepSeek-V4.1-Flash completes 43.3% of tasks without breaking a rule, against 64.0% for the top model tested, Claude Opus 5.

Read More: the previous open-source radar, on MiMo-V2.6, Kandinsky 6.0 Video and EmbeddingGemma 2 and the September radar, on the earlier DeepSeek-V4.1-Flash release

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Open {Models} {This} {Month:} {DeepSeek’s} {Update,} a 1 {GB}
    {Assistant} {Model,} and a {Reader} for {100-Page} {Documents}},
  date = {2026-11-02},
  url = {https://toknow.ai/posts/open-source-radar-deepseek-1017-lfm3-unlimited-ocr-2-nemotron-transcribe/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Open Models This Month: DeepSeek’s Update, a 1 GB Assistant Model, and a Reader for 100-Page Documents.” https://toknow.ai/posts/open-source-radar-deepseek-1017-lfm3-unlimited-ocr-2-nemotron-transcribe/.