DeepSeek releases V4 Flash Vision, an open MIT-licensed 305B VLM
Permissive weights bolt vision onto DeepSeek's V4 agent stack — read charts, screens and PDFs. Also inside: Grok Bot's $20 route, and AMD's local-AI tower.
Copy markdown
MIT weights, no strings attached
DeepSeek-V4-Flash-Vision-Exp is a 305B mixture-of-experts model built on the V4-Flash base and shipped under a plain MIT license — commercial use, fine-tuning and redistribution are all fair game. The safetensors are on Hugging Face now.
The unlock: agents that can see
It bolts visual understanding onto DeepSeek's agent stack while holding text-agent parity, so you can point it at screenshots, charts and PDFs to build screen-driving or document-reading agents this weekend. It posts 83.9 on Terminal Bench 2.1 and 64.3 on Chartography.
'Exp' means verify before you ship
The 'Exp' tag is literal: the model is experimental and its benchmarks are self-reported, with thin outside evaluation so far. Treat headline numbers like ZeroBench 35.0 as vendor claims until independent runs confirm them.
Elsewhere: Grok Bot's $20 on-ramp
xAI's always-on Grok Bot agents, once gated behind $200-plus tiers, now unlock on the $20/mo Cursor Pro plan, with Android joining Mac, Windows and iOS. The catch: xAI still won't publish usage allowances, and heavy users report a weekly quota gone in a day.
Elsewhere: AMD's trillion-param desktop
AMD unveiled the Threadripper Halo Station at IFA — 96 cores and dual MI350P accelerators with up to 576GB of HBM3e, pitched to run trillion-parameter models locally. But at a reported $100K-$150K and shipping in 2027, it's a lab flex, not a home rig.