Breaking

DeepSeek Releases Open-Weight V4 Models, Drawing Microsoft's Interest but Still Trailing Top U.S. Systems

July 21, 2026 at 11:01 EDT

Foundation Models
Open Source
Infra & Chips

Chinese AI lab DeepSeek has released its DeepSeek-V4 series as open-weight models, winning praise from researchers and prompting Microsoft to host the systems on Azure while it weighs a secured version for its Copilot agent features—yet experts caution the models still lag the best proprietary U.S. offerings in important respects.

April 24, 2026 · DeepSeek

DeepSeek-V4 Goes Open-Weight — and Microsoft Puts It on Azure

A pair of open Mixture-of-Experts models wins praise for coding and long context at a fraction of proprietary prices — even as experts warn they still trail the best U.S. systems on hard reasoning.

1.6T

Total params in V4-Pro (~49B active/token)

284B

Total params in lighter V4-Flash (~13B active)

Token context — roughly 15–20 novels of text

V4-Pro benchmark scores

Competitive on standard tests; reasoning mode reported near Claude Opus-class on SWE-Bench Verified

90.1

MMLU

76.8

HumanEval

73.5

MMLU-Pro

Scores on a 0–100 scale.

The cost edge driving adoption

On Microsoft Foundry, V4-Pro input pricing runs roughly $1.74–$1.93 per million tokens — the kind of low-cost open model hyperscalers see as a way to blunt rising inference bills as Copilot usage climbs.

Why developers are warm

Strong coding & agentic performance
1M-token long-context tasks
Open weights — run & fine-tune locally
Tight Azure integration for enterprises

Why caution remains

Still trails top U.S. proprietary systems
Gaps on complex reasoning & knowledge
Security concern given Chinese origin
Microsoft weighs a data-isolated deployment

The open question

Enterprise uptake of the V4 series will turn on one thing: whether the cost edge can be captured without ceding control over sensitive data — which is why Microsoft's interest hinges on a "secured" deployment that keeps enterprise data off servers in China.

Continue reading

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