DeepSeek released the weights and inference code for V4-Flash-Vision-Exp on Hugging Face on August 31 under an MIT license, marking the company's first open-source multimodal model2. The 305-billion-parameter model had been available through the DeepSeek API since August 21 and reached third-party gateways such as OpenRouter on August 271,5.
The release adds image understanding to DeepSeek's budget V4 Flash text model, accepting JPEG, PNG, GIF, and WebP inputs for tasks including chart reading, screenshot OCR, and visual question answering. DeepSeek flagged the weights as an experimental build, and the vision-exp release serves as a validation vehicle for architecture choices, including the aligner, DSpark, and Hyper-Connections components, before a broader multimodal release.
The open-source package includes model files, a tokenizer, a prompt-encoding reference implementation, and a minimal PyTorch inference stack covering the vision encoder, an aligner, DFlash attention, a mixture-of-experts routing layer, Hyper-Connections, and DSpark modules.
On pure-text tasks such as agents, reasoning, and world knowledge, DeepSeek says performance matches the release version of V4-Flash. On agent benchmarks requiring visual understanding, the company says the model gains substantially over V4-Flash and brings multimodal agent capability close to Claude Opus 4.8. Tech Times noted that these benchmark claims still need independent verification4.
The model's pricing sets up a direct comparison with Google's Gemini 3.7 Flash, released August 13. V4-Flash-Vision-Exp is priced at $0.22 per million input tokens and $0.66 per million output tokens, with prices doubling during weekday peak hours. Gemini 3.7 Flash bills at $0.75 and $3.75 per million on OpenRouter.
A head-to-head test by The New Stack ran both models through three image tasks mimicking back-office work: chart reading, invoice auditing, and incident diagnosis from production logs. Both models achieved 9/9 accuracy across the test suite. On the chart test, both correctly identified that operating costs exceeded total revenue in Q1, named Subscriptions as the revenue segment that grew every quarter, and estimated full-year revenue at $36.1 million. On the invoice audit, both caught a line-total error where 10 units at $45.99 were printed as $505.89 instead of $459.90, calculated the correct total due as $3,958.89, and flagged a due date (July 28) that preceded the invoice date (August 12). DeepSeek additionally noted that the printed subtotal did not match the printed line items. On the incident-diagnosis test, both models pinpointed 14:05:12 as the time the outage began and named the batch job as the root cause. Gemini went further by suggesting specific PostgreSQL commands to kill the batch job.
The accuracy parity came with a sharp cost-speed tradeoff. DeepSeek averaged 16.8 seconds per response at a total test cost of $0.0039, while Gemini averaged 7.2 seconds at $0.0122. Token economics differed as well: DeepSeek counted each image at roughly 500 prompt tokens versus Gemini's roughly 1,150. On the invoice test specifically, DeepSeek took 30.5 seconds and billed 3,467 completion tokens; Gemini answered in 7.9 seconds with 944 completion tokens.
ANALYSIS The matched accuracy at roughly one-third the cost positions V4-Flash-Vision-Exp as a viable alternative for latency-tolerant, cost-sensitive document-processing pipelines, while Gemini 3.7 Flash retains an advantage where response time matters. DeepSeek advises testing the model against specific workloads rather than assuming production-grade behavior across all image tasks.