DeepSeek Just Dropped a 305B Vision Monster and Left Silicon Valley Sweating Again
Because nothing says 'we are just getting started' quite like dropping a massive multimodal model with open weights just ten days after the last routine update.
DeepSeek has officially published the weights for DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model packed with 305 billion parameters distributed under a generous MIT License. The architecture builds directly upon the recent V4-Flash release by bolting on a visual module and rigorous multimodal training to handle screenshots, dense charts, and complex agentic tool use.
On benchmark tracks like ApexBench, the new model pulls in 36.5 points compared to its text-only predecessor at 26.2, while simultaneously edging out Opus 4.8 on tasks like DeepSWE with 59.3 points against 58.0. Text processing and tool invocation have also seen steady bumps, with Toolathlon-Verified climbing to 75.9, though niche edge cases like CyberGym saw a slight dip down to 75.3.
The accompanying repository includes everything required for independent deployment, featuring a dedicated tokenizer, minimal PyTorch inference scripts, visual encoders, MoE blocks, and Hyper-Connections. Developers can interface with the system using vLLM, SGLang, Transformers, or Docker containers, while the open-source community has already begun pushing out localized quantized builds.
Silicon Valley budget planners are likely pouring strong coffee this morning as open-weights engineering casually redefines what constitutes a mid-cycle experimental checkpoint.
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.