2026-07-26

Datalab rewrote Marker as a three-mode pipeline. Version 2 hits 76.0 on olmOCR-bench and sustains 2.9 pages per second on one B200 — over 5× MinerU's pipeline backend, while beating Docling on both accuracy and speed. Here's how it compares against MinerU, Docling and LiteParse, and which one fits your use case. The post Datalab Marker v2 vs MinerU, Docling, and Liteparse: Benchmark Breakdown appeared first on MarkTechPost .

单卡突破10000 video tokens/s，Vivix打通实时多模态生成全链路

Article URL: Comments URL: Points: 32 # Comments: 0

Give Claude Code and Codex a voice Discussion | Link

Please be honest. I would love to hear about guys really dedicated to local AI and who really reject subscriptions (especially to openai and anthropic). What do you use your model for?

My goal was to create a benchmark to measure the spatial awareness and memory of models. Eventually, I came up with the simple idea of a maze where the model must find a key and use it to open the escape door. Here’s the difference to a normal maze, however! The model CANNOT see the whole map. At each step, it only gets feedback on its immediate surroundings within the overall maze. Thus, in order to succeed, it must be able to track its position and orientation within the coordinate space. Even

It randomly shut off on me for about 15 seconds a few days ago, killing power to my PC and server. After it turned back on the estimated runtime dropped to only 5 minutes, at a supposed full charge, but wasn't telling me to replace the batteries. I knew they were around 5 years old so I immediately ordered replacements. Once I pulled out the old ones I was met with this beauty. I'm glad I didn't hesitate in ordering their replacements.

OpenAI's fancy new AI keypad will be a lot of fun for some, while many others are probably not going to touch it.

Please let me know which small LLM model you're using and what you're using it for.

Since a lot more people are trying to build their own multi-GPU machines, I thought I should help to prevent a common mistake people make with building multi-GPU machines. Which is using an Intel consumer platform like Z890 for multi-GPU setups. Although the CPU provides 24 PCIe 5.0 lanes with 16x available to bifurcate to 8x8x on two PCIe x16 slots on the higher end boards, this is completely useless for AI inference/training workloads that require P2P between the GPUs. In my testing I used an

So, this came up on: Archived version is case it disappears: