CONFIRMED: DeepSeek open-sources V4-Flash-Vision-Exp

Editorial label: CONFIRMED. Basis: the official deepseek-ai Hugging Face repository, with weights, tokenizer, a minimal inference implementation and an MIT license. DeepSeek-V4-Flash-Vision-Exp is the first experimental multimodal model in the V4 family. It has 305B parameters, builds on V4-Flash-0731 and adds a vision encoder plus an Aligner. Primary source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

What we know

The official card says multimodal agent scores rise substantially while text-only agent scores stay comparable. DeepSeek's own table: ApexBench Pass@1 moves from 26.2 to 36.5 (Opus 4.8 at 39.4); Agents' Last Exam lands at 27.3, ahead of Opus 4.8's 25.7; DeepSWE reaches 59.3 versus Opus 58.0. NL2Repo still trails, 57.7 versus 69.7. The product intent is an agent that reads screenshots, UIs and charts and then calls tools, not showcase VQA. The repo includes vLLM and SGLang recipes with DSpark. Secondary reports say the API arrived around August 21 and the weights about ten days later.

What we still do not know

There is no long blog post or vision paper on this page. Vision API pricing, multimodal context length and Portuguese quality are not on the card. The benchmarks are DeepSeek's, not an independent harness. The Exp tag still means experimental.

Why it matters

Teams building screen agents, visual QA or dashboard readers get an MIT 305B open-weight model in the Opus 4.8 neighborhood on some multimodal agent tests, without waiting on a closed lab. It is not an automatic replacement for Fable 5.1 or Muse Spark: it is the first set of eyes on the V4 line, cheap to license and heavy to host. Test it on product screenshots and tool use; do not migrate production on the card alone.