DeepSeek launches experimental vision version of Flash model, now available on OpenRouter

DeepSeek has released DeepSeek V4 Flash Vision Exp, an experimental variant of its Flash model that adds image-processing capability without changing the pricing: $0.22 per million input tokens and $0.66 per million output tokens (roughly 0.81 and 2.44 shekels respectively). The model, an experimental build of V4 Flash 0731, ships with a one-million-token context window and is immediately accessible through the OpenRouter API.

Under the hood sits a sparse MoE (mixture-of-experts) architecture with 284 billion total parameters, of which only 13 billion are active on each forward pass. That design keeps cost low and latency low while layering multimodal abilities on top of the existing text foundation. DeepSeek positions the model for document and chart understanding, VQA (visual question answering), and agent workflows that interleave text and images.

According to the manufacturer, the new model matches V4 Flash 0731 on text-agent benchmarks, reasoning, and world-knowledge tests. The claimed edge appears in two specific evaluations — Agents' Last Exam and ZeroBench — where it reportedly outperforms Claude Opus-4.8. The caveat is significant: these are vendor-supplied figures only, with no independent verification or third-party runs under identical conditions. The Experimental tag means a future stable release could alter behavior or performance.

OpenRouter reports a P50 (50th percentile) throughput of 113 tokens per second and a median latency of 1.03 seconds, the best numbers among providers currently serving the model. Uptime measures 100% and 99.96% on two separate metrics, backed by automatic failover to an alternate provider when the primary errors. Those figures reflect OpenRouter's serving infrastructure, not the model in isolation, and apply only to customers consuming the service through OpenRouter.

What remains absent is a technical report, open weights, and results on standard vision benchmarks such as MMMU or MathVista. The pricing mirrors the text-only model, making experimentation relatively cheap, but the lack of independent benchmarks means any production integration today is a calculated gamble. For developers already building on OpenRouter who want to test multimodality without switching providers, this is a convenient entry point. Everyone else will wait for the non-experimental release and independent data.