DeepSeek V4 Arrives: 1.6 Trillion Parameters, $0.28 Output Tokens, and a Direct Challenge to the West
China's DeepSeek has released a preview of its long-awaited V4 model in two variants: a flagship Pro model with 1.6 trillion total parameters and a lean Flash variant priced at just $0.28 per million output tokens. The release comes more than a year after R1 shook global markets, and signals that Chinese AI development has not slowed.
More than a year after DeepSeek’s R1 reasoning model rocked global tech markets, the Chinese AI startup has returned with its long-awaited V4 model. DeepSeek V4, released in preview on April 24, 2026, arrives in two variants and reasserts the company’s position at or near the frontier of large language model development.
The release carries competitive weight far beyond its benchmarks. In a global AI race where the United States has imposed sweeping chip export restrictions on China and European governments are rethinking AI sovereignty, DeepSeek’s continued progress on a fraction of Western compute budgets is a direct challenge to the assumption that frontier AI requires frontier hardware spending.
Two Models, One Architecture
DeepSeek V4 ships as two distinct models. V4-Pro is the flagship: a 1.6 trillion total-parameter Mixture of Experts (MoE) architecture with 49 billion parameters active per inference pass. V4-Flash is a smaller, faster variant at 284 billion total parameters with 13 billion active, designed for latency-sensitive and cost-sensitive applications.
Both models share a 1,048,576-token context window — one million tokens — allowing entire codebases, lengthy legal contracts, or multi-volume research corpora to be processed in a single prompt. According to MIT Technology Review, that context length matches what is offered by cutting-edge versions of models like Gemini and Claude.
The headline architectural innovation is a new attention design that combines token-wise compression with DeepSeek Sparse Attention (DSA). Traditional transformer attention scales poorly across extremely long contexts, creating a well-known performance degradation problem. The design compresses older information and focuses on the parts most likely to matter, while keeping nearby text in full so it does not miss important details. The result, according to DeepSeek, is drastically reduced compute and memory costs for long context: in a 1-million-token context, V4-Pro uses only 27% of the computing power of V3.2 and cuts memory use to 10%.
Pricing That Resets Expectations
DeepSeek has again used pricing as a weapon. V4-Pro costs $3.48 per million output tokens via the DeepSeek API. V4-Flash costs just $0.28 per million output tokens. For comparison, OpenAI charges about $30 per million output tokens for comparable work. Anthropic charges about $25 per million output tokens for the same amount of work.
At $0.28 per million tokens, V4-Flash is cheaper than sending a postcard. For developers building high-throughput applications — customer service bots, document processors, code review pipelines — the economics are transformative. Even V4-Pro at $3.48 is far cheaper than the $30 and $25 that OpenAI and Anthropic charge for the same amount of work.
DeepSeek says the new attention design sharply reduces the compute and memory cost of long context, and it expects to lower V4-Pro prices later this year as Huawei scales up Ascend 950 production. The company has not disclosed its training costs for V4, but its prior models have been trained for a fraction of what comparable U.S. labs spend.
Open Source and Huawei Integration
Both V4 variants are released as open-weight models, available for download and fine-tuning on Hugging Face. This continues DeepSeek’s tradition of open releases that have repeatedly seeded new research and derivative models across the global AI community.
More geopolitically notable is the model’s close integration with Huawei’s Ascend AI chips. U.S. export controls bar Chinese developers from buying Nvidia’s most advanced AI chips. Rather than stalling, DeepSeek appears to have adapted — optimizing V4 for Huawei’s domestic alternative. Huawei said its Ascend supernode products, based on the Ascend 950 series, would support DeepSeek V4.
This detail matters beyond DeepSeek itself. It demonstrates that U.S. chip restrictions, while impactful, have not succeeded in preventing frontier model development in China — they have instead accelerated domestic chip optimization.
Benchmark Performance
DeepSeek has positioned V4-Pro as near-frontier on coding, mathematics, and reasoning tasks. On benchmarks shared by the company, DeepSeek says V4-Pro competes with leading closed-source models such as Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro, though its tech report says V4 falls marginally short of GPT-5.4 and Gemini 3.1 Pro. Independent evaluators on platforms like Chatbot Arena have begun placing V4-Pro in the top five globally, though comprehensive third-party benchmark results are still emerging.
V4-Flash, despite its dramatically smaller active parameter count, reportedly performs comparably to GPT-4o-mini and Gemini 3.1 Flash on most standard evaluations — a ratio of price-to-performance that will be difficult for Western providers to match.
The Geopolitical Subtext
This release lands in a charged context. Chinese regulators have recently been scrutinizing foreign investment in domestic AI companies including Moonshot AI, StepFun, and ByteDance. Meanwhile, U.S. policymakers are debating further tightening of the chip export regime. DeepSeek operates in this environment as something of an anomaly: a Chinese AI lab that remains largely open, internationally accessible, and technologically credible.
When DeepSeek R1 dropped in January 2025, the immediate assumption in Western AI circles was that it represented a one-time disruption — clever engineering of an existing paradigm, not a sustained capability trajectory. V4’s release challenges that assumption. The company has shipped V2, V3, R1, and now V4 in roughly eighteen months, each meaningfully more capable than the last.
What This Means for Enterprise Adopters
For CTOs and AI leads evaluating model providers, V4’s release opens practical questions. V4-Flash’s price point makes it compelling for any application currently using GPT-4o-mini or Gemini Flash — the cost savings are too large to ignore. V4-Pro enters serious consideration for workloads where model quality is paramount and data residency requirements permit use of DeepSeek’s API.
The persistent concern is reliability and enterprise support. DeepSeek’s API infrastructure, while significantly more robust than it was a year ago, still has occasional availability issues under peak demand. For mission-critical enterprise deployments, many organizations will want to run open-weight versions on their own infrastructure — which the open-source release makes possible.
The message from Hangzhou is clear: the AI race remains wide open, the cost curve is still dropping, and Western providers cannot take their lead for granted.
Correction (November 15, 2026): 14 sentences were corrected or removed to match the sources listed on this page.