DeepSeek V4-Flash-Vision-Exp: Multimodal Flash Model With Near Opus 4.8 Claims
The release extends DeepSeek's Flash tier with visual-agent capabilities at a price point below its own R1-0528 model, potentially lowering costs for workflows requiring repeated visual input. DeepSeek's benchmark claims against Opus 4.8 remain unvalidated by independent real-world testing.
DeepSeek released V4-Flash-Vision-Exp on August 21, 2026, on its API platform, adding multimodal input to the Flash model line while retaining V4-Flash's text capabilities, including agents, reasoning, and world knowledge. DeepSeek stated the model makes a major leap over V4-Flash on multimodal agent benchmarks, bringing performance close to Opus-4.8. iweaver.ai noted that in DeepSeek's published benchmarks, V4-Flash-Vision-Exp wins some tests while Opus 4.8 wins others, and that independent real-world comparisons have not yet been conducted to validate those claims. One account reported the model outperforms a previous July checkpoint on most benchmarks. DataLearnerAI published scores of 83.90 on Terminal-Bench 2.1 (11th of 44 models), 59.30 on DeepSWE (12th of 27), 57.70 on NL2Repo-Bench (3rd of 8), 64.30 on Chartography, 36.50 on ApexBench, and 27.30 on Agents' Last Exam, all in MaxTools mode with thinking enabled. A single image input is capped at 384 tokens and billed at standard V4-Flash text rates. V4-Flash-Vision-Exp is priced below DeepSeek-R1-0528, and the two models share no common benchmark datasets, making direct comparison between them difficult. DeepSeek Harness 0.1.1, released the same day, includes support for the new model.
A threads.com account, reported August 28, said V4-Flash-Vision-Exp outperforms DeepSeek's own July checkpoint on most benchmarks and characterized the August 21 API release as made without formal announcement beyond a repo update.
Transformer News separately affirmed the claim that the model approaches Opus 4.8 on agent benchmarks, without adding new data.
Transformer News affirmed DeepSeek's claim that the model rivals Opus 4.8 multimodal agentic capabilities
DeepSeek released V4-Flash-Vision-Exp on August 21, adding multimodal input to V4-Flash and claiming performance close to Opus-4.8 on agent benchmarks.
DataLearnerAI published scores of 83.90 on Terminal-Bench 2.1 and 59.30 on DeepSWE, run in MaxTools mode with thinking enabled. iweaver.ai noted independent real-world comparisons have not yet validated those claims, and that image inputs cap at 384 tokens billed at V4-Flash text rates. llm-stats.com added V4-Flash-Vision-Exp shares no benchmark datasets with DeepSeek-R1-0528, limiting cross-model comparison.
The Neuron Daily covered the release, noting it maintained the Flash tier's low-cost positioning
DeepSeek released V4-Flash-Vision-Exp on its API platform; DeepSeek Harness 0.1.1 released same day with support for the model
I send a short email each morning with the stories that moved. If you would rather just read here, that works too.
Subscribe free