DeepSeek-V4-Flash

Claims and evidence

• Released 24 April 2026 ● Primary source[1] [3]Date: 2026-04-24 [4]General Information: Release date

• Status AvailablePreview weights remain downloadable from the deepseek-ai Hugging Face organisation. In the DeepSeek API the preview was replaced by DeepSeek-V4-Flash-0731 under the unchanged model name deepseek-v4-flash on 2026-07-31. ● Primary source[2]Model Downloads; License [3]Date: 2026-07-31
• Replaced by DeepSeek-V4-Flash-0731In the DeepSeek API: DeepSeek-V4-Flash-0731 took over the model name deepseek-v4-flash on 2026-07-31; its card says it supersedes the preview. ● Primary source[3]Date: 2026-07-31 [5]Introduction

• Successor of DeepSeek-V3.2 EditorialEditorial link: from 2026-04-24 the legacy API names deepseek-chat and deepseek-reasoner, which had served DeepSeek-V3.2 since 2025-12-01, pointed to the non-thinking and thinking modes of deepseek-v4-flash; the card also compares base results with DeepSeek-V3.2. No source calls V4-Flash the successor of V3.2. ● Primary source[3]Date: 2026-04-24; Date: 2025-12-01 [2]Base Model evaluation

• Change · Architecture Uses the V4 hybrid attention design (Compressed Sparse Attention plus Heavily Compressed Attention), Manifold-Constrained Hyper-Connections and the Muon optimizer.Compared with DeepSeek-V3.2 ● Primary source[2]Introduction
• Change · Size 284B total and 13B activated parameters, against 671B total and 37B activated for DeepSeek-V3.2-Base in the card's base-model table.Compared with DeepSeek-V3.2 ● Primary source[2]Base Model evaluation
• Change · Context length Supports a 1M-token context, which DeepSeek made the default across its official services with V4.Compared with DeepSeek-V3.2 ● Primary source[1]
• Change · Availability Took over the legacy API names deepseek-chat and deepseek-reasoner from DeepSeek-V3.2, with those names scheduled to be discontinued on 2026-07-24.Compared with DeepSeek-V3.2 ● Primary source[3]Date: 2026-04-24; Date: 2025-12-01

• Input text ● Primary source[4]Model properties: Modalities
• Output text ● Primary source[4]Model properties: Modalities
• Feature Reasoning modeNon-think, Think High and Think Max modes; the API offers thinking and non-thinking modes. ● Primary source[2]Instruct Model: reasoning effort modes [1]API is Available Today [4]Model properties: Modalities
• Feature Long context ● Primary source[1] [2]Introduction
• Open weights YesWeights and code under the MIT License. ● Primary source[2]License [4]Methods of distribution and licenses [1]
• Context window 1M tokens ● Primary source[2]Model Downloads [1] [4]Model properties: Context length
• Parameters 284B total / 13B activeMixture-of-Experts: 284B total parameters, 13B activated per token. The V4 model card PDF on the Transparency Center gives 285B total parameters instead. ● Primary source[1] [2]Model Downloads [6]Abstract
• Access API, consumer app, Open-weights download ● Primary source[1] [4]Methods of distribution and licenses

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from DeepSeek on AI Radar →