DeepSeek open-sources a 305B multimodal model, ten days after quietly launching it as a paid API

DeepSeek released the weights for DeepSeek-V4-Flash-Vision-Exp on Hugging Face on August 31, under an MIT license, ten days after the same model quietly went live on DeepSeek's own API. It's the first experimental multimodal model in the V4 family: a 305-billion-parameter mixture-of-experts build that adds a vision encoder and aligner to DeepSeek-V4-Flash-0731's existing architecture, continuing training so the model can actually reason about images rather than ignore them. DeepSeek's own model card says the base V4-Flash-0731 text model 'ignores the multimodal elements in the input' on certain benchmarks -- a plain admission that visual understanding is new ground here, not a rebrand of an existing capability. The model carries the same 1,048,576-token context window as its text-only sibling.

The gap between API and open weights is DeepSeek's usual pattern, not a new caution: a paid window first, then a full release once the model has run in production. Ten days is short by that standard, and shorter than it looks next to rival Z.ai, which released its comparably sized GLM-5.3-Flash with weights available immediately rather than staged behind an API at all. Both approaches now land in roughly the same place -- an MIT-licensed, 300-billion-class multimodal MoE anyone can download -- which makes the gap between them a question of sequencing, not of whether either company actually intends to open-source. On DeepSeek's API, the model prices at $0.22 per million input tokens and $0.66 per million output, with cached reads at $0.007 per million; on Hugging Face, it ships with reference support for vLLM (with speculative decoding and FP8 KV-cache) and SGLang, plus eight ready-made quantizations for llama.cpp, LM Studio, Jan and Ollama.