DeepSeek v4: the most expected open-source model ever released, and the quietest landing

2026-05-22 · Source: The Lambda Deep Learning Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Advanced, short

Summary

DeepSeek has released version 4 of its open-source models, DeepSeek V4 Pro and DeepSeek V4 Flash, under an MIT license, marking a quiet but significant landing after 15 months of anticipation. DeepSeek V4 Pro features 1.6T parameters, making it the largest open-source model to date, deployable on a single NVIDIA HGX B200 node. The smaller DeepSeek V4 Flash has 284B parameters. The architecture emphasizes serving efficiency over benchmark scores, introducing a hybrid attention mechanism with Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). This innovation reduces single-token inference FLOPs and KV-cache memory at long context, leading to a reported 10x drop in cost-to-serve and 10x less memory compared to V3.2, aiming for a reliable 1M context window. While these engineering wins are substantial for cost, the model's long-context benchmark gains and overall evaluation scores against competitors like Kimi K2.6 have been mixed.

Key takeaway

For MLOps Engineers managing large language model deployments, DeepSeek v4 offers significant cost-to-serve reductions, particularly for long-context workloads. If your applications prioritize a 1M context window and operational efficiency, you should evaluate V4 Pro or Flash for potential 10x inference cost savings. Be aware that raw benchmark scores are mixed, so validate performance against your specific use cases rather than relying solely on general evaluations.

Key insights

DeepSeek v4 prioritizes serving cost reduction through architectural innovation over raw benchmark performance.

Principles

• Serving cost optimization drives architectural design.
• Hybrid attention can significantly cut inference FLOPs.
• Open-source models can achieve 1M context windows.

Method

DeepSeek V4 employs a hybrid attention mechanism, extending DeepSeek Sparse Attention with Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to reduce inference FLOPs and KV-cache memory.

In practice

• Deploy DeepSeek V4 Pro on NVIDIA HGX B200 for scale.
• Evaluate V4 for workloads valuing context over raw evals.
• Consider V4 Flash for smaller, competitive deployments.

Topics

• DeepSeek V4
• Large Language Models
• Open-source AI
• Inference Optimization
• Hybrid Attention
• Cost-to-Serve

Best for: NLP Engineer, CTO, AI Architect, AI Engineer, MLOps Engineer, Machine Learning Engineer

Related on AIssential

• DeepSeek V4 - almost on the frontier, a fraction of the price — DeepSeek V4 models offer competitive performance and industry-leading low costs through significant efficiency gains.
• DeepSeek-V4: The Most Powerful Open-Source Model Ever — DeepSeek-V4 commoditizes high-reasoning AI with open-source models, massive context, and disruptive pricing.
• Deepseek v4: Best Opensource Model Ever? (Fully Tested) — Deepseek v4 models offer cost-efficiency and long context, but real-world performance may not match benchmark claims.
• Three reasons why DeepSeek’s new model matters — DeepSeek V4 offers competitive performance and cost-efficiency with a 1-million-token context window, optimized for Chinese chips.
• DeepSeek V4: Does China Just Matched GPT-5? — DeepSeek V4 offers near-frontier performance at significantly lower cost and open weights, challenging established proprietary models.
• DeepSeek V4: One Million Tokens, Three Thinking Modes, and the First Real Hands-On Reports — DeepSeek-V4 optimizes long-context LLMs via novel attention and training stability, offering tiered reasoning and cost-effective variants.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Lambda Deep Learning Blog.