Breaking

DeepSeek Pushes New V4-Pro Model, $52 Billion Valuation and Custom Chips as It Chases US Rivals

July 17, 2026 at 11:42 EDT

Foundation Models
Open Source
Funding & M&A

DeepSeek, the Hangzhou-based startup that jolted Silicon Valley last year, is now pursuing several ambitious efforts at once—flagship model releases, its first outside fundraising and an IPO push, and in-house chip development—even as the performance gap with America's top AI labs shows signs of widening.

April 24, 2026 · DeepSeek

DeepSeek's V4 Bet: Cheaper Frontier AI, Three Races at Once

The Hangzhou lab ships its open-weight V4-Pro and faster V4-Flash while chasing its first fundraise, an IPO, and in-house chips — even as the gap with US frontier labs shows signs of widening.

The core pitch — price

API cost vs. comparable Western models

10×

Western
offerings

1×

DeepSeek
API

DeepSeek's API runs about one-fifth to one-tenth the cost of comparable Western offerings — developers call it a "price revolution."

1.6T

total parameters (MoE) — only ~49B active per token

80.6%

SWE-Bench Verified — closing in on Claude Opus

token context window · 32T+ pretraining tokens

Three races at once

$52B

valuation in first external round — ~$7B raised, Tencent among backers

IPO

preparing a mainland-China listing; weighing more fundraising

Chips

building in-house inference silicon to cut Nvidia & Huawei reliance

The upside

Open-weight, MIT license — self-host with no vendor lock-in
Led 5 of 7 benchmarks in NIST's CAISI cost-efficiency test
Top-ranked on independent coding benchmarks; strong long-context

The caveats

Chinese frontier models trail top US systems by ~8 months
Beaten on some benchmarks by GLM and Kimi despite huge size
Preview instability, peak-time latency, domestic-hardware bottleneck

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.