DeepSeek V4 Has Launched: Open-Sourced on April 24, 2026 — Specs and Highlights Recap

The wait is over. DeepSeek V4 officially launched on April 24, 2026, fully open-sourced under the MIT license, with model weights published on Hugging Face. All the earlier speculation about V4's timing — TechNode reports, HuggingFace upload activity, and competitive pressure from GPT-5.4 and Claude 4.6 — has now been settled. This article recaps the core facts, specifications, and benchmark results of the release.

Release at a Glance

• Release date: April 24, 2026 — DeepSeek-V4 preview officially launched.
• License: MIT, with weights published on Hugging Face.
• Two editions: DeepSeek-V4-Pro and DeepSeek-V4-Flash.
• Core focus: Agentic coding + 1M-token context + extreme efficiency.
• Legacy retirement: deepseek-chat and deepseek-reasoner will be retired on July 24, 2026.

Two Editions, Detailed Specs

DeepSeek shipped two editions at once, covering everything from high-end reasoning to low-cost, high-throughput workloads:

Architecture Highlight: Hybrid Attention (CSA + HCA)

V4's real efficiency breakthrough lies in its MoE (Mixture of Experts) + hybrid attention architecture, which combines:

• CSA (Compressed Sparse Attention)
• HCA (Heavily Compressed Attention)

This design means that at a 1M-token context, per-token compute is roughly 27% of V3.2's, and KV Cache memory is about 10% of V3.2's. In other words, V4 turns ultra-long context from an "expensive luxury" into an affordable default — and that is precisely why V4 can offer million-token context at such low prices.

Benchmarks (V4-Pro, Measured)

V4-Pro delivers top-tier results among open-source models across major benchmarks:

The SWE-bench Verified score of 80.6% is the highest among current open-source models, putting V4 firmly in the top tier for agentic coding.

API Pricing

After a 75% price cut at launch, V4's long-term pricing is highly competitive:

Compared to closed-source frontier models, V4 is roughly 5–30x cheaper while being fully open source — a combination no closed-source competitor can match.

How to Use DeepSeek V4

V4 is live now. Here's how to access it:

Option 1: Web at chat.deepseek.com

Head to chat.deepseek.com to try it directly, with two modes:

• Expert Mode: For complex reasoning and agentic coding tasks.
• Instant Mode: For faster, lower-cost everyday queries.

Option 2: Official API

Sign up at platform.deepseek.com, create an API key, and call the model deepseek-v4-pro.

Note: the legacy models deepseek-chat and deepseek-reasoner will be retired on July 24, 2026, so plan your migration to V4 soon.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Implement quicksort in Python"}],
)
print(response.choices[0].message.content)

Option 3: Atlas Cloud and Local Deployment

• Atlas Cloud: A DeepSeek-recommended partner offering hosted access at launch — atlascloud.ai.
• Local deployment: V4 is open-sourced under MIT, with weights on Hugging Face. You can self-host via inference frameworks such as vLLM and TGI.

A Look Back at Market Expectations

Before V4's official launch, the industry pieced together its timing from several independent signals: TechNode's reporting on internal progress, dense upload activity on HuggingFace, and the competitive pressure created by the back-to-back launches of GPT-5.4, Claude 4.6, and Gemini 3.1 Pro — all pointing to an imminent next-generation model from DeepSeek. Those expectations have now been met: V4 is officially released, answering the market with efficiency and open-source commitment that exceeded what many predicted.

Conclusion

DeepSeek V4's launch brings together five defining traits: agentic coding ability (SWE-bench 80.6%), a 1M-token context window, CSA+HCA extreme efficiency, full open source (MIT), and very low pricing. For developers, now is the best time to get hands-on: try it at chat.deepseek.com, or plug the API into your workflow.

Sources

The following is a summary of DeepSeek's official release (2026-04-24):

• DeepSeek official launch announcement (April 24, 2026)
• Hugging Face DeepSeek organization (model weights)
• chat.deepseek.com and platform.deepseek.com

Some third-party benchmark figures may change as evaluations are updated.

Last updated: April 24, 2026