DeepSeek V4 Drops With 1.6T Parameters, Bargain Pricing, and Huawei Chips: OpenAI Should Pay Attention

Exactly one year after R1 sent shockwaves through Wall Street, DeepSeek is back with a sequel. The Chinese lab officially launched the open-source preview of DeepSeek-V4 on April 24, and the specs are wild.

DeepSeek V4 Pro and Flash: two models, one million tokens

V4-Pro packs 1.6 trillion total parameters with 49 billion active, while V4-Flash is the leaner sibling at 284 billion total and 13 billion active. Both support a native one-million-token context window, a massive leap from the 128,000 tokens of V3.

The secret sauce? A hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). In the 1M-token setting, V4-Pro requires only 27% of V3.2's inference FLOPs and 10% of the KV cache. That is a staggering efficiency gain.

Near-frontier performance at bargain-bin pricing

V4-Flash costs just $0.14 per million input tokens, undercutting GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku. V4-Pro is priced at $0.145 input / $3.48 output, cheaper than Gemini 3.1 Pro, GPT-5.5, and Claude Opus 4.7. DeepSeek is basically daring you not to try it.

The lab's own tech report admits V4 trails GPT-5.4 and Gemini 3.1 Pro by roughly three to six months on knowledge benchmarks. But on coding, V4-Pro-Max hits 93.5 on LiveCodeBench, a new open-model record. For teams that care about code and agents more than trivia, that trade-off may be easy to accept.

Huawei chips and the geopolitics of AI silicon

Huawei confirmed on Friday that its Ascend AI processors can support DeepSeek's V4 model. Whether V4 was actually trained on Huawei silicon remains unclear, but the inference story is real. Counterpoint analyst Wei Sun said V4's ability to run natively on local chips could help Beijing achieve greater AI sovereignty.

Chinese chip manufacturers surged in Hong Kong trading, with SMIC up 9% and Hua Hong Semiconductor up 15%. Still, Morningstar analyst Ivan Su told CNBC that V4 is unlikely to match R1's market shock because traders already priced in competitive Chinese AI. The surprise factor is gone; the pressure is not.

What DeepSeek V4 means for developers

V4 is optimized for popular agent tools such as Anthropic's Claude Code and OpenClaw. The API supports both OpenAI ChatCompletions and Anthropic formats, so migration is nearly painless. If you're building agentic workflows on a budget, DeepSeek just handed you a very compelling option, and OpenAI's freshly launched GPT-5.5 will need more than brand prestige to justify the price gap.