Featured
▲ Datalab released OmniParseBench, an open OCR benchmark with 16,288 pass/fail tests to audit document parsers. Each test is 1 yes/no question about 1 page. 2,937 pages from 2,343 documents, 92 languages, 22 suites. A parser's score is its pass rate. The headline is the mean of text, table, and layout; tags let you slice results but never change a verdict. On the launch board, Gemini 3.8 Flash leads at 92.9. Datalab's accurate mode is second at 92.7. The top 3 sit within 1.3 points. Claude Sonnet 5.5 tops text (93.1) and tables (95.4) but lands at 88.7 because layout is 77.5. Tesseract scores 17.6 on text and 0.0 on tables. 2 asterisks matter: Gemini 3.8 Flash and Sonnet 5.5 helped create or verify some tests. Tables are 6,767 tests; English has 10,089 tests to 1,931 for all other languages. Datalab built the benchmark and sells a parser, and says so. Its advice is to run your own documents. pip install omni-parse-bench
Also this week’s main signals
▲ SpaceXAI released Grok Voice 3.0, a voice agent model that ranks #1 on Artificial Analysis. Speech in, speech out, with reasoning and tool calls in between. It scores 86.9 on the AA Speech-to-Speech Index. GPT-Live-1 (Astra, medium) is at 83.2, Gemini 3.8 Live Extended Thinking at 82.6.
▲ 🐟 🐟 Free to read and Pay when it acts. 30% extra on every top-up through Oct 31 2026. Claim 30% Bonus [Sponsored]
▲ Fish Audio Releases Drama 3, a 35B Open-Weight MoE Text-to-Speech Model A TTS model that takes plain-language acting direction, up to 16 speakers per request. 35B total, about 3B active: 8 of 256 experts plus 1 shared. Fish Audio reports 93.2% on AA Pronunciation Robustness, #1, up from 88.1% for S2.1 Pro.
▲ TypeSafe AI Releases Jev 1.14 With Images, Box Type and 262k Context Live as jev-preview; jev-latest switches in about a week. 3 experimental features. Images: PNG or JPEG, 256 to 1,568 px per side, gated behind a "danger zone" header. Box: 1 normalized bounding box per question, edges 0 to 1. It always returns a box, even when the object is absent, so pair it with a Noul check. In TypeSafe's internal test, 4 per-edge Noul questions caught 30 of 36 cut-off crops. Long context: 262K tokens opt-in, default 64K.
Other Signals
▲ Modal's Series D takes the serverless GPU company from $4.65B to a reported $15B valuation.
▲ Character.AI filed Chapter 11; Disney's $20M stalking-horse bid buys a non-exclusive IP license, not the app.
▲ 🐟 🐟 Free to read and Pay when it acts. 30% extra on every top-up through Oct 31 2026. Claim 30% Bonus [Sponsored]
▲ NVIDIA's cuPotential is a CUDA-X library for agents that simulate atomic and molecular interactions.
▲ OpenCode 2.0 runs terminal, web, and desktop clients on 1 background service, with background subagents and async shell.
▲ openKylin's AI-era Linux plan moves the OS toward "intent management" on its KART runtime. RVV kernels: 2 to 5× on FP32.[Sponsored]
▲ Chai Discovery signed GSK after GSK's wet lab found binders to every target tested; terms undisclosed.
▲ CopilotKit's CLiF turns agent traces plus user actions into skills, evals, and fine-tuning data. Early access, self-hostable.[Sponsored]
Until next week,
The Marktechpost Team
Sponsorship Opportunities