DeepSeek V4 Flash: Small Model, Big Ambitions

DeepSeek released a preview of two new-generation models at the end of April 2026 — DeepSeek V4 Pro and DeepSeek V4 Flash. Both are built on the Mixture of Experts (MoE) architecture, which activates only a portion of parameters for each query, thereby dramatically reducing operating costs. While V4 Pro is the flagship with 1.6 trillion parameters (49 billion active), it is V4 Flash that has attracted the most attention — it has 284 billion total parameters, but only 13 billion active, yet its performance approaches models that cost dozens of times more.

Both models handle a context window of 1 million tokens — that's roughly 750,000 words, equivalent to the entire Lord of the Rings trilogy. For developers, this means the ability to insert an entire code repository or extensive documentation into a single prompt.

Price Shock: How Much Each Model Costs

DeepSeek's biggest weapon is not performance, but price. Let's look at a price comparison per million tokens (input/output):

• DeepSeek V4 Flash: $0.14 / $0.28
• DeepSeek V4 Pro: $1.74 / $3.48
• GPT-5.5: $5 / $30
• Claude Opus 4.7: $5 / $25
• Claude Sonnet 4.6: $3 / $15
• Gemini 3.1 Pro: $5 / $25

In practice, this means that V4 Flash is 35× cheaper on input and 107× cheaper on output than OpenAI's GPT-5.5. For a company that processes tens of millions of tokens through models monthly, the difference amounts to hundreds to thousands of dollars per month. It's no wonder that, according to Ramp's data, DeepSeek became the fastest-growing software vendor among American companies in June 2026.

How V4 Flash Performs in Benchmarks

DeepSeek itself admits that V4 Pro lags behind the best Western models by 3 to 6 months — specifically behind GPT-5.4 and Gemini 3.1 Pro in knowledge tests. In programming benchmarks, however, both V4 models are "comparable to GPT-5.4". On Artificial Analysis's GDPval-AA benchmark, V4 Pro achieved 1,554 Elo points, a jump of 355 points compared to the previous V3.2.

The Flash variant is not designed as a "lightweight" version, but as a standalone model optimized for agent tasks — i.e., for autonomous AI agents that independently perform tasks, search for information, or work with code. DeepSeek states that V4 models are integrated with tools such as Claude Code, OpenClaw, and OpenCode and are used internally for agent programming.

Technical Magic: How DeepSeek Achieved Such Efficiency

A key innovation is a new hybrid attention architecture that combines token compression with so-called "sparse attention". The result? According to the technical report, V4 Pro requires only 27% of the computational power and 10% of the KV cache memory compared to the older V3.2 when processing a million tokens. The Flash version goes even further — 10% of computations and 7% of cache.

In translation: DeepSeek has proven that even huge contexts can be processed with minimal overhead. This is crucial for agentic AI, which often has to maintain long conversational histories or work with extensive datasets.

The models were trained on up to 33 trillion tokens with an emphasis on multilingual data, scientific publications, and agent scenarios. DeepSeek explicitly states in the technical report that during post-training, it uses distillation from its own specialized models — teachers for mathematics, code, agents, and instruction following. Accusations of distillation from GPT or Claude, which Anthropic and OpenAI raised against DeepSeek, are neither confirmed nor explicitly refuted in the technical report.

Chinese Price War: DeepSeek, Xiaomi, and Alibaba Drive Prices Down

DeepSeek is not the only Chinese player pushing AI prices to a minimum. In May 2026, the company announced that it was making the 75% discount on the V4 Pro model permanent — output tokens thus cost at least 34× less than with GPT-5.5. Meanwhile, Xiaomi with its MiMo model discounted its API by 99%, and Alibaba with Qwen3.6-27B beats GPT and Gemini in programming benchmarks for a fraction of the price. Smaller players are also emerging — China's Z.ai with the GLM-5.2 model, which surpasses GPT-5.5 in programming at one-sixth the cost.

This price war also impacts Western companies. OpenAI has admitted that current prices are not sustainable in the long term and has introduced a lower ChatGPT Go tariff for 120 CZK per month. Microsoft is reportedly limiting the use of Claude Code within the company due to rising costs. And Uber, according to reports, exhausted its annual budget for AI tools in just four months.

What This Means for Czechia and Europe

For Czech companies and developers, DeepSeek V4 Flash is an extremely interesting alternative. The model is available as open-weight under the MIT license — meaning you can download it for free, run it on your own infrastructure, and even use it commercially. For a startup or a medium-sized company that doesn't want to pay hundreds of dollars a month for APIs from OpenAI or Anthropic, this is an attractive path.

The DeepSeek API is accessible via an interface compatible with both OpenAI and Anthropic formats, so transitioning from ChatGPT or Claude does not require significant code rewriting. The model supports Czech — although it primarily targets English and Chinese, its multilingual training on 33 trillion tokens also includes Slavic languages.

However, there's also a downside. Data sent to the DeepSeek API travels to Chinese servers, which carries security and legal risks, especially in light of GDPR and the upcoming EU AI Act. If a company processes sensitive data, it is more reasonable to run the model locally or through European inference providers (for example, Fireworks AI or DeepInfra), who also offer V4 models.

DeepSeek on the Rise: American Companies Choose Cheaper AI

Data from the payment platform Ramp, which tracks transactions of over 50,000 American companies, show that DeepSeek led the "trending" category in June 2026 — meaning suppliers with the largest relative growth. This is not the first time Chinese models have seen massive adoption: already in December 2025, Chinese models for the first time surpassed American ones in the number of downloads on Hugging Face, where they accounted for over 44% of all downloads of popular new models.

Ramp's chief economist Ara Kharazian warns, however: "American companies pay DeepSeek directly and send it their data. This is not about the benefits of open-source." He also doubts that this trend will last — especially if regulatory pressure between the US and China intensifies.

And it is already intensifying. In April 2026, the US accused China of industrial theft of intellectual property from American AI labs. The context is also important for European companies: while DeepSeek is freely available today, geopolitical tensions can change the rules of the game at any time — from API blocking to a ban on using Chinese models in certain sectors.