DeepSeek open-sources V4-Flash-Vision: MIT multimodal weights, 305B
The API launched Aug 21; today the weights land free on Hugging Face — native vision, 1M-token context, $0.22/M in off-peak, but not yet a clean local run.
Copy markdown
Open weights, MIT, native vision
DeepSeek dropped V4-Flash-Vision-Exp's weights on Hugging Face today under the MIT license — its first multimodal model, with vision built into the architecture (a native DFlash visual path) instead of a bolted-on tower like V3.2. That means commercial fine-tuning and self-hosting with no strings.
1M context, images at 384 tokens each
On the hosted API it takes mixed text+image (base64, URL, or Files API) across a 1,048,576-token window, up to 600 images per request billed at just 384 tokens each. Off-peak runs $0.22/M input ($0.007 on a cache hit) and $0.66/M output; the 01:00–10:00 UTC peak window doubles both.
Local runs work — but not drop-in
The 305B-parameter MoE (the V4-Flash base activates ~13B per token) fits a single DGX Spark / GB10 desktop, but early testers say you need vLLM plus vision patches rather than a clean load, and ~4 hours to pull the weights over a 100 Mb/s line. Text-only stacks run now; vision tooling is still catching up.
Edges Opus 4.8 on vision, trails on text
DeepSeek's self-reported numbers put it at 35.0 on ZeroBench (Pass@5), just past Opus 4.8's 34.0, with 83.9 on Terminal-Bench 2.1 and 36.5 on ApexBench — while trailing Opus by 1–12 points across text-agent tasks. No third-party verification yet, so treat the vision wins as provisional.