# DeepSeek V4

> **TL;DR.** DeepSeek V4 is a series of open-weight Mixture-of-Experts language models from DeepSeek that bring one-million-token context to frontier-class performance at a fraction of closed-source pricing.

- **Category:** AI / LLM / Open-Weight Models
- **Stage:** validating
- **Age:** 53 days
- **Origin date:** 2026-04-24
- **First detected:** 2026-04-24
- **Canonical URL:** https://earlyterms.com/term/deepseek-v4
- **Sources:** 8 primary URLs

## Definition

DeepSeek V4 is a series of open-weight Mixture-of-Experts language models from DeepSeek that bring one-million-token context to frontier-class performance at a fraction of closed-source pricing. The series ships as two variants: V4-Pro (1.6T parameters, 49B active) and V4-Flash (284B parameters, 13B active).

Released on [April 24, 2026](https://api-docs.deepseek.com/news/news260424), both models feature a novel hybrid attention architecture — Compressed Sparse Attention plus Heavily Compressed Attention — that cuts KV-cache requirements to 10% of DeepSeek V3.2's at 1M-token context. Weights are MIT-licensed on Hugging Face; the API is live and supports both OpenAI ChatCompletions and Anthropic formats.

## Analogy

Think of it as a turbocharged open-source engine for long-context AI workloads at cloud-compute prices.

## Why it's emerging now

DeepSeek shipped V4 on April 24, 2026 — a same-day open-weight release with 1M-token context and pricing that undercuts every comparable closed model. V4-Flash at $0.14/MTok is cheaper than OpenAI's cheapest frontier option; V4-Pro at $1.74/MTok is the most affordable premium-tier model available.

## Related terms

- *parent:* DeepSeek V3.2
- *related:* DeepSeek-R1
- *competitor:* qwen3
- *competitor:* claude-opus-4-7
- *competitor:* claude-opus-4-6
- *related:* openclaw
- *related:* managed-agents
- *related:* agentic-coding
- *related:* context-window
- *related:* agent-harness
- *related:* deep-research
- *parent:* Mixture of Experts

## Sources

1. [DeepSeek — V4 Preview Release announcement](https://api-docs.deepseek.com/news/news260424)
2. [DeepSeek-V4-Pro — Hugging Face model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)
3. [Simon Willison — DeepSeek V4 analysis](https://simonwillison.net/2026/Apr/24/deepseek-v4/)
4. [TechCrunch — DeepSeek previews V4](https://techcrunch.com/2026/04/24/deepseek-previews-new-ai-model-that-closes-the-gap-with-frontier-models/)
5. [South China Morning Post — Huawei full support for V4](https://www.scmp.com/tech/big-tech/article/3351239/deepseek-releases-next-gen-ai-model-world-leading-efficiency)
6. [Hacker News — DeepSeek v4 thread (1,372 pts)](https://news.ycombinator.com/item?id=47884971)
7. [vLLM Blog — DeepSeek V4 efficient long-context attention](https://vllm.ai/blog/deepseek-v4)
8. [CNBC — DeepSeek V4 preview, AI race intensifies](https://www.cnbc.com/2026/04/24/deepseek-v4-llm-preview-open-source-ai-competition-china.html)

---
_Generated by EarlyTerms · https://earlyterms.com/term/deepseek-v4_