DeepSeek-V4 - The open-source era of 1M context intelligence
DeepSeek-V4 Preview is a new series of highly efficient MoE language models, featuring V4-Pro (1.6T params) and V4-Flash (284B params). Both models support a 1 million token context window by default, utilizing a novel hybrid attention architecture to drastically reduce compute and memory costs.
Replies
Flowtica Scribe
Hi everyone!
The long-awaited DeepSeek V4 is finally here, and the message is simple: 1M context is becoming normal.
V4-Pro is the flagship model, with stronger agentic coding, world knowledge, and reasoning. V4-Flash is the fast, efficient version for more economical use. Both models support 1M context and are available through API today, with open weights already released.
DeepSeek’s real ambition here is to make frontier long-context intelligence more accessible, just like it has been doing all along🫡
P.S. Think about all the quota and money you’ve burned through just to unlock massive context windows in Codex or CC. Well, let’s look forward to a future where that no longer feels like a luxury. Thanks, DS!💙
open weights + 1M context is a serious shift — most "long context" models choke on retrieval at the tail. how does V4-Flash hold up on needle-in-a-haystack at 800k+? curious if Engram memory actually moves the needle there.
fwd2cal
@zaczuo amazing work to you and the team. I can’t believe this is open weights.