DeepSeek open-weights its first vision model: MIT, 1M context
284B params, 13B active, 1M context, MIT — and it edges Opus-4.8 on agent benchmarks. The catch: a 168GB checkpoint, so 'local' means a multi-GPU box.
Copy markdown
MIT weights, self-host and sell it
DeepSeek released V4-Flash-Vision-Exp under a plain MIT license — 284B total params, 13B active per token, a 1M-token native context, and a vision encoder trained onto the V4-Flash base. You can fine-tune it, self-host it, and ship it inside a commercial product with no strings attached.
It edges Opus-4.8 on agent tasks
Vendor benchmarks put it at 27.3 on Agents' Last Exam — just past Opus-4.8's 25.7 — plus 59.3 on DeepSWE, while the vision graft lifts ApexBench to 36.5 from the text-only 26.2. The text and agent skills survived the multimodal training intact.
'Local' comes with a 168GB asterisk
The FP8/FP4 checkpoint is roughly 168GB, so this is multi-GPU-server territory, not a laptop. But NVFP4 and Unsloth quants are already up and DGX Spark owners are running it, and it ships with SGLang plus speculative decoding to cut inference cost.
What you could ship this weekend
With open MIT weights you can stand up a self-hosted visual agent — screenshot-to-action, diagram and document understanding, UI regression QA — on hardware you control. On DeepSeek's hosted API, a new Files API lets you upload an image once and reuse it free across calls, with images billed at up to 384 tokens each.