DeepSeek Opens Up a 305 Billion Parameter Model That Can Finally See
DeepSeek-V4-Flash-Vision-Exp went up on Hugging Face under an MIT license: the first V4 family model that handles images, with big jumps on multimodal agent benchmarks and text performance that holds steady.
DeepSeek published a new model on Hugging Face without much noise around it. DeepSeek-V4-Flash-Vision-Exp is the first model in the V4 family that can look at pictures as well as read text, it weighs in at 305 billion parameters, and it ships under an MIT license, which is about as permissive as software licensing gets. The company describes it as experimental, and the name carries the “Exp” to prove it.
The approach is straightforward: take the existing V4-Flash architecture, bolt on vision modules, and keep training. What comes out is a model that got substantially better at multimodal agent work while text-only performance stayed roughly where it was. On ApexBench, a multimodal agent test, it scores 36.5 against 26.2 for the previous V4-Flash-0731, though that comparison is generous, since the older model simply ignores images in the input. On Agents’ Last Exam it posts 27.3, edging past the 25.7 that DeepSeek’s own table lists for Opus 4.8. Text-side numbers are close to unchanged: 83.9 on Terminal Bench 2.1, 59.3 on DeepSWE. The repository ships a tokenizer, a reference implementation, and serving recipes for vLLM and SGLang, and quantized community versions appeared almost immediately.
Two terms worth unpacking. Parameters are the adjustable numbers inside a model, the things training tunes. 305 billion of them is large, and it means this is not something you run on a laptop. Open weights means DeepSeek published those numbers for anyone to download, which is different from open source software: you get the finished model, not the recipe or the training data. The MIT license is the notable part, because it puts almost no restrictions on commercial use. A European company can download this, run it entirely on its own hardware, and never send a single customer document to anyone else’s server. That is a real option that did not exist a few years ago, and it is why open weight releases from Chinese labs keep mattering to businesses that have nothing to do with China.
What this means for you. If you use AI through a chat window, nothing here is for you today, and that is fine. If you build things, this is worth a look: a permissively licensed model that reads charts, screenshots and documents, at a size you can actually self-host if you have the hardware, is a useful thing to have. Worth keeping expectations grounded, though. “Experimental” in a model name usually means rough edges, the benchmark comparisons come from DeepSeek rather than an independent evaluator, and the honest reading of the numbers is that this matches the top proprietary models on some multimodal agent tasks and trails them on others. The direction is what is interesting, not the leaderboard position.
Sources
- DeepSeek-V4-Flash-Vision-Exp model card (Hugging Face)
- DeepSeek-V4 collection (Hugging Face)
Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
The EU Just Classified ChatGPT as a Search Engine, and That Changes What It Owes You
The European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act, alongside Reddit and Roblox as platforms. All three cleared 45 million monthly EU users and have until January 2027 to comply.
