DeepSeekV4

Twitter 2026-04 technology active Updated 2026-05-31
Mid 2020s

First documented in April 2026 on Twitter. Currently active and in regular use across social platforms.

Also known as: DeepSeekV4ProDeepSeekV4FlashDeepSeekPro

#DeepSeekV4 is the hashtag for the Chinese AI lab’s April 24, 2026 dual-model release — V4-Pro and V4-Flash — both shipped as open weights under MIT license on Hugging Face. With V4-Pro scoring 80.6% on SWE-bench Verified at roughly one-sixth the per-token cost of OpenAI’s GPT-5.5, the launch repeated and amplified the cost-shock dynamic that DeepSeek’s January 2025 R1 model first set off, and re-opened the “open-source frontier” debate that had dominated AI Twitter through early 2026.

Quick Facts

  • Release date: April 24, 2026
  • Models: DeepSeek V4-Pro (1.6T parameters MoE, 49B active) and V4-Flash (284B parameters, 13B active)
  • License: MIT (open weights, free commercial use)
  • Context window: 1 million tokens; max 384K output tokens
  • Headline benchmark: 80.6% on SWE-bench Verified (V4-Pro-Max) — highest at release
  • Cost vs GPT-5.5: roughly $348 vs $3,000 per 100M output tokens

What It Is

DeepSeek V4 is a dual-model release built on a Mixture-of-Experts (MoE) architecture. V4-Pro is a 1.6-trillion-parameter MoE with 49B active parameters, designed for complex reasoning, coding, and agentic tasks. V4-Flash is a lighter 284B-parameter model (13B active), optimized for speed and cost — trailing V4-Pro by just 1.6 points on SWE-Bench while costing roughly 25× less per output token. Both ship under MIT license with open weights on Hugging Face, available via the DeepSeek API on day one.

The architecture introduces a hybrid CSA+HCA attention mechanism that cuts FLOPs to about 27% and KV cache to roughly 10% of V3.2, along with a Muon optimizer — the technical details that explain how the headline cost numbers are even possible.

Why It Mattered

At release, V4-Pro-Max posted 80.6% on SWE-bench Verified and 93.5 on LiveCodeBench — the highest coding-benchmark score of any model at the time. Independent reviewers calculated that for applications producing roughly 100 million output tokens per month, DeepSeek V4 ran around $348 versus about $3,000 for GPT-5.5 — a roughly 1/6th-the-cost story that travelled fast on tech Twitter.

This is the second time DeepSeek has shocked the market in 15 months. #DeepSeek’s R1 release in January 2025 had already triggered a panic in tech stocks — Nvidia lost hundreds of billions in market cap in a single day — and resurfaced debates about export controls and the geopolitics of AI. V4 made it clear that wasn’t a one-off.

Cultural Impact

On the AI-policy conversation. DeepSeek V4 landed inside a busy first week of May 2026 that also saw OpenAI’s GPT-5.5 Instant and Google’s Gemma 4. TechRadar’s weekly roundup called the release one of the most surprising developments of the week, framing it as evidence that open-weight models are “closing the capability gap” with closed frontier labs.

On developer adoption. Within weeks, integration guides (“multi-model routing with V4 + GPT-5.5”) and side-by-side benchmark posts dominated developer-facing newsletters and YouTube channels. Coverage in DataCamp, Codersera, and MindStudio framed V4 as the first credible open-source frontier-class model.

On the geopolitics of AI. CNN’s coverage on the day of release explicitly asked whether V4 would “make waves like last year” — and within a month, the answer was clearly yes. DeepSeek had gained roughly 75 million downloads in January 2025 alone after R1; V4 extended that user base into a broader developer audience and gave open-weight advocates a concrete data point.

  • #DeepSeek — the umbrella tag for the company and its earlier R1 release
  • #OpenSourceAI — the broader policy/capability debate V4 sits inside
  • #GPT55 — the closed-API contender released the same week

Sources

Frequently Asked Questions

What is #DeepSeekV4? +

DeepSeek V4 is a pair of frontier-class language models released on April 24, 2026 by Chinese AI lab DeepSeek — V4-Pro (1.6 trillion parameter MoE, 49B active) for complex reasoning and coding, and V4-Flash (284B parameter, 13B active) optimized for speed and cost. Both ship as open weights on Hugging Face under MIT license.

How does DeepSeek V4 compare to GPT-5.5? +

On the SWE-bench Verified coding benchmark, V4-Pro scores 80.6% — the highest coding benchmark score of any model at release. For applications producing roughly 100 million output tokens per month, the per-token cost works out to about $348 with DeepSeek V4 versus about $3,000 with GPT-5.5.

Why did DeepSeek V4 reignite the AI-cost debate? +

Because it pairs frontier-level benchmarks with an MIT-licensed open-weights release at a fraction of closed-API pricing. That combination repeats and amplifies the dynamic DeepSeek's January 2025 R1 release first set off — pressuring closed-API economics and reigniting the geopolitics-of-AI conversation.

Sources & References

Explore #DeepSeekV4

Related Hashtags

2010 2026 #DeepSeekV4 2026 #AI 2010 #FOTM 2011 #PodcastSpeed 2015 #LLM 2023 #LLaMA 2023 #DeepSeek 2025
Related hashtags by year of first appearance — circle size reflects lifetime volume, fade reflects how active each tag still is.