DeepSeek V4 Flash and V4 Pro ship with 1.6T-param agentic-coding tier

DeepSeek V4 is the company's first explicit two-tier release, abandoning the single-base-model approach of V3 in favor of the Flash/Pro split that Anthropic (Claude Haiku/Sonnet/Opus), OpenAI (mini/standard/pro), and Google (Flash/Pro) have all adopted. V4 Flash is the high-throughput MoE designed for batch inference, agentic loops with many low-stakes tool calls, and cost-sensitive pipelines. V4 Pro, at 1.6T total parameters, targets the opposite end: long-horizon reasoning, complex agentic coding, and analysis tasks where token cost matters less than completion quality.

Mechanically, both models inherit DeepSeek's signature MoE architecture and aggressive training-cost discipline. The company hasn't published full benchmark detail at launch, but historically DeepSeek releases hit within 5–10% of frontier US labs at a fraction of the price — a dynamic that has reshaped pricing pressure across the entire industry over the last 18 months.

Competitive context: V4 lands in a week dominated by Google I/O and OpenAI cloud-restructuring news, which may dilute attention but also reflects DeepSeek's positioning — it competes on price-per-token and developer trust, not on PR cycles. Alibaba's parallel Qwen3.7 preview release (Max-Preview at LM Arena rank 13 for text, Plus-Preview at rank 16 for vision) shows the Chinese frontier remains active even as US policy attention is focused elsewhere.

Skeptical takes: 1.6T total parameters in V4 Pro raises inference-cost questions for self-hosters, and the Flash/Pro split risks fragmenting the developer experience that made DeepSeek attractive in the first place. Watch for: independent benchmark numbers in the next two weeks, hosting availability on Together / Fireworks / Hyperbolic, and whether NVIDIA releases an NVFP4-quantized version (as it did this week for Moonshot's Kimi-K2.6) to ease GPU deployment.