DeepSeek V4 models

6 versions from DeepSeek, reviewed by the TopReviewed AI panel.

ModelScoreStatusReleasedPrice in/outContext
DeepSeek V4-Flash
frontier-adjacent quality at the lowest cost in market
8.7preview2026-04-23$0.14 / $0.281MReview →
DeepSeek V4-Pro
frontier-grade agentic coding at open-weights cost
8.5preview2026-04-23$0.43 / $0.871MReview →
DeepSeek V3.2
open-weights math/reasoning at GA stability
8.2GA2025-11-30$0.25 / $0.38128KReview →
DeepSeek V3.1
the model that put DeepSeek on the production-agent map
7.7GA2025-08-20$0.21 / $0.79128KReview →
DeepSeek R1 (0528)
exposed-CoT reasoning at a fraction of o-series cost
8.0GA2025-05-27$0.55 / $2.19128KReview →
DeepSeek-VL2
open-weights OCR and document understanding at the edge
7.2GA2024-12-12 / 4KReview →

Strongest: DeepSeek V4-Flash

DeepSeek V4-Flash is the volume tier of the V4 family — a 284B-parameter Mixture-of-Experts model that activates just 13B parameters per token, inherits V4-Pro's 1M-token context and CSA/HCA sparse attention, and serves it at $0.14 in / $0.28 out per 1M tokens. It shipped as a preview on 2026-04-24 with MIT-licensed open weights, and it is the default model behind the legacy `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) aliases, which retire 2026-07-24. The single sentence a buyer needs: for high-throughput chat, agent inner loops, and batch processing where per-token cost is the binding constraint, V4-Flash is the cheapest frontier-adjacent model in the market by a wide margin. - **Provider:** DeepSeek - **Released:** 2026-04-24 (preview) - **Status:** Preview (no GA announced) - **Context window:** 1,000,000 tokens - **Max output:** 384,000 tokens - **Modalities:** Text in / text out - **Knowledge cutoff:** 2026-02 - **Headline price:** $0.14 in / $0.28 out per 1M tokens

Full review →