DeepSeek V4-Flash-Vision Open-Source Hands-On: An 8GB Vision Module Gives the Model Eyes
DeepSeek open-sources its first multimodal model V4-Flash-Vision with only 8GB of added vision weights at no extra API cost.
DeepSeek has open-sourced the full weights (168GB FP8) of V4-Flash-Vision-EXP on HuggingFace, its first multimodal model. Built on the V4-Flash-0731 text foundation with an ~8GB vision plugin module, it's priced identically to the text-only version with the same time-based discount strategy. Local deployment is currently limited as vLLM and SGLang haven't completed support yet, but developers have already validated its value via API — using screenshot feedback loops to build vision Agent workflows and complete multi-capability platforms integrating TTS, OCR, and voice cloning in a single day. Across small model comparisons, the V4-Flash series remains the top choice for cost-efficient development.
DeepSeek V4-Flash-Vision Goes Open Source: First Multimodal Model Weights Released
DeepSeekAd has made another move. The company recently open-sourced the full weights of its first multimodal model, V4-Flash-Vision-EXP, on HuggingFace — marking a significant milestone for this cost-efficiency-focused Chinese AI lab, which has now officially added "vision" to its lineup. The model's API had already quietly launched on the official website, but now that the weights are public, developers can finally attempt local deployment and build their own vision Agent applications.
What you might not have noticed is that this isn't an entirely new architecture trained from scratch. Instead, it's built on top of DeepSeek V4-Flash-0731, with a multimodal plugin module attached. In other words, the underlying text capabilities are inherited directly from the original Flash model, and the visual capability is "stacked on top" via an additional module.
In terms of size, V4-Flash-Vision weighs in at approximately 168GB (FP8 weights), compared to 160GB for the original V4-Flash. That means the entire multimodal vision capability adds only about 8GB of weights. This detail is quite telling — the fact that 8GB is enough to give a model "eyes" suggests that multimodal implementation may not be as complex as many assume. It's largely about efficient visual alignment on top of a mature text foundation.
Timeline: From May Beta to Official Open Source
DeepSeek began testing its vision capabilities as early as May of this year, though initially only through the web interface. An API version followed later, and now the weights have been officially open-sourced. This pacing is somewhat out of character for DeepSeek, which typically follows a "release and open-source simultaneously" approach.
The model architecture is identical to V4-Flash-0731, yet the weights weren't immediately released. The exact reasons are unknown — it may relate to multimodal module integration, safety reviews, or other engineering considerations. This phased "API first, open source later" strategy reflects the additional complexity involved in shipping multimodal models.
Local Deployment Status: vLLM and SGLang Support Not Yet Available
Developers looking to deploy V4-Flash-Vision locally will need to wait a bit longer.
One notable anomaly: DeepSeek has not yet provided vLLM deployment commands, and mainstream inference frameworks vLLM and SGLang have not yet completed support for this model. Typically, when major labs release a model, inference framework teams announce compatibility on the same day — for instance, some Qwen models already had official compatibility confirmed at launch. Some labs even hand models to inference framework teams in advance, so developers can sometimes detect upcoming models through early code changes in inference engines.
DeepSeek took a different approach this time. Since the model upload was only just completed, mainstream frameworks like vLLM and SGLang haven't yet added support for the Vision-EXP multimodal extension module. While they previously supported the text-only V4-Flash, the new vision module will take some additional time. Support announcements are expected soon, but local deployment carries a meaningful barrier in the short term.
Hands-On Testing: Building a Multimodal App Platform in a Day
Beyond the parameter-level discussion, real-world projects have validated the model's development efficiency. According to one tester, his model-serving website was built almost entirely using DeepSeek V4-Flash-Vision, with no other models involved.
The platform not only provides an API interface but also integrates several local model capabilities — including TTS (text-to-speech), speech recognition, OCR, voice cloning, and image captioning — along with simplified demo apps for each feature to verify functionality and connect to the API.
When questioned whether the core development was really done in a single day, the tester confirmed it was. The approach involved running three or four sub-Agents in parallel to develop different modules simultaneously. Crucially, the model's vision capability allowed developers to share screenshots of the actual rendered frontend, enabling rapid identification and resolution of errors. This "screenshot → analyze → adjust prompt → regenerate" feedback loop is precisely the core advantage multimodal vision Agents have over pure text models.
DeepSeek's Time-Based Pricing Strategy: Cost Efficiency Is the Key Selling Point
On pricing, V4-Flash-Vision's API is priced identically to V4-Flash — more capability at no extra cost. Even more notable is DeepSeek's time-based pricing strategy:
- Weekends: Half the standard price
- Off-hours (after work): Half the weekday price
This is one reason many developers choose to do intensive development on weekends. Based on real-world usage feedback, this is considered one of the "best value-for-money multimodal models" currently available.
Head-to-Head Comparison: Who Wins Among Small Models?
Compared to similar small models, the results are fairly clear-cut:
- Gemma 5.3 Flash: Poor user experience — frequent interruptions and stuttering have led many developers to remove it from their regular toolkit. Even being free isn't enough to meet real development needs. Performance on OpenRouter may differ from the official version, and inference stability is questionable.
- Qwen3-8B Flash: Performance is "not far behind" DeepSeek's 27B model. It has a native context window of 256K, extendable to 1M, with a 1-million-context API already available on the official website, making it competitive for long-context scenarios.
Overall, the DeepSeek V4-Flash series remains the cost-efficiency benchmark in the small model space. The addition of vision capability only further cements its lead in lightweight, high-value development scenarios.
Conclusion: What DeepSeek's Multimodal Open Source Means
The open-sourcing of DeepSeek V4-Flash-Vision represents a meaningful step forward for Chinese large language models in the multimodal space. The minimal 8GB vision module addition, pricing parity with the text-only version, and the demonstrated visual Agent development capabilities in real-world testing all make it highly competitive on a value-per-dollar basis.
That said, local deployment support still requires waiting for vLLM and SGLang compatibility. For cost-conscious developers, this is a model worth watching closely — especially for building automated development workflows that require "visual understanding." The actual local inference performance and the timeline for framework support remain to be further validated.
Related articles
LLM Application Development: A High-Paying Career Guide for Ordinary People
LLM application development offers fresh graduates up to 35K/month, with SP offers reaching 600K. This article explains the core skills, how to break in, and how ordinary people can seize AI opportunities through learning by doing.
Say Goodbye to AI Influencer Faces: A Complete Guide to Creating Visual Assets for AI Comic Dramas from Scratch
Bilibili creator shares a full workflow for zero-basics AI comic drama visual assets: use Doubao Work skills with script-first, visual references, and real-person texture to escape AI influencer faces, then lock character and scene consistency with asset sheet layouts, weapon detail images, scene nine-grids, and cinematic color palettes.
The AI Production Pipeline Behind Great Videos: From Research to Agent Validation
AI video production isn't just a model race. This article breaks down a transferable four-step pipeline: researching professional workflows, organizing a skill checklist, configuring agents, and experimental validation—and explains how general-purpose LLMs like DeepSeek can replace paid tools to run the entire process.