April 2026 · AI News

DeepSeek V4 Launch — V4-Pro and V4-Flash Released Simultaneously,
Open Weights Up to 30x Cheaper Than GPT-5.5

On April 24, 2026, DeepSeek released V4-Pro and V4-Flash at the same time. After shaking up the industry once with V3, they dropped two models in a single shot. Both are MIT-licensed open weights.

Bottom line first. V4-Pro sits between GPT-5.2 and GPT-5.4 on reasoning benchmarks. Coding performance hits a Codeforces rating of 3,206 — 23rd among all registered human users worldwide. The API price is about 8.6x cheaper than GPT-5.5 on output tokens. V4-Flash is even more extreme. It's roughly 107x cheaper than GPT-5.5 on output.

The timing is what makes this launch stand out. Right after GPT-5.5 captured the industry's attention with top-tier performance, DeepSeek fired back with "comparable performance, drastically lower cost." The fact that it's open weights — meaning you can run it on your own servers without paying API fees — adds even more weight. This article breaks down both models in numbers: architecture, performance, pricing, what the MIT license actually means, and how to call the API directly.

V4-Pro and V4-Flash — Architecture and Specs

Both models use a MoE (Mixture of Experts) architecture. Think of MoE as a massive factory with hundreds of specialized production lines — only the lines relevant to each incoming job get activated. A fraction of the total parameters handles any given inference. Compute cost drops dramatically relative to total parameter count.

V4-Pro has 1.6T total parameters, but only 49B are active during inference. V4-Flash has 284B total with 13B active. The numbers make V4-Pro look enormous, but the actual inference cost is determined by active parameters. Both models include the Hybrid Attention Architecture and support up to 1M tokens of context.

DeepSeek releasing both models at the same time was deliberate positioning. V4-Pro prioritizes performance. V4-Flash prioritizes speed and cost. Instead of one model trying to cover everything, they split the roles from the start. It's the same strategy Anthropic uses with Opus, Sonnet, and Haiku.

One key architectural feature is Multi-Head Latent Attention (MLA). It compresses the KV cache from standard transformers to reduce memory usage. First introduced in V3, it was refined further in V4. Both inference speed and cost improvements trace back to this design.

How MoE Architecture Lowers Costs

MoE isn't a simple numbers game. V4-Pro has 1.6T parameters, but only 49B activate per inference. The remaining 1.55T are dormant for that pass. Inactive parameters don't participate in computation. Power consumption and GPU memory usage are both governed by active parameters.

A gating network decides which Experts get activated. The gating network acts as a lightweight classifier. For each token, it selects the best combination of Experts. Based on the V3 architecture, only a small number of Experts activate per token out of hundreds available. This uses far less compute than a dense model.

The effect is even more pronounced in V4-Flash. 13B active out of 284B total is about 4.6%. V4-Pro is about 3.1% — 49B active out of 1.6T total. That's why the API is cheap. The number of parameters actually doing math is far lower than a dense model of equivalent size.

MoE has a drawback. You need all parameters loaded into GPU memory for self-hosting, which means VRAM requirements are large. Loading V4-Pro in FP16 theoretically requires about 3,200GB of VRAM. When using the API, this doesn't matter — but if you're planning to self-host, quantization is mandatory.

Hybrid Attention — What the 1M Token Context Actually Means

Attention is how an AI model computes relationships between words in a sequence. Think of it like reading an entire thick novel while keeping front-to-back context in memory at once. The problem is that standard Full Attention scales quadratically as token count grows. Processing 1M tokens with Full Attention explodes in compute cost.

DeepSeek V4's Hybrid Attention Architecture solved this. It mixes Full Attention with efficient Linear Attention. Sections requiring precise attention use Full Attention; the rest use Linear Attention. The result is compute cost that scales closer to linearly for 1M token inputs.

To put 1M tokens in perspective: one English novel is roughly 100K tokens. 1M tokens equals about 10 novels. For a large codebase, that's hundreds of thousands of lines of code. The effective character count will be lower in practice since some languages tokenize less efficiently than English.

In practice, use cases that actually need 1M context are limited. It's useful in agent systems that need to reference multiple documents simultaneously. Legal and financial domains processing hundreds of pages of contracts in a single pass qualify too. For a standard chat application, 128K is more than enough. The value of 1M context is the design flexibility of having no ceiling — not that you should always use it.

Codeforces 3,206 — A New Benchmark for Coding Performance

Codeforces is a competitive programming platform. A rating of 1,200 is beginner level. 1,900 is Expert. 2,400 and above is Grandmaster. 3,206 is above that. DeepSeek V4-Pro hit this score, placing it 23rd among all registered human users worldwide.

This number measures pure algorithmic problem-solving ability. It covers dynamic programming, graph traversal, mathematical optimization, and complex data structure design. It isn't a perfect equivalent to writing production code. But it's a strong enough signal that "this AI has the logical reasoning capacity of a top 0.001% human programmer."

The improvement in AI coding performance over time has been steep. When GPT-4o launched, its Codeforces rating was in the 1,100s. In under two years, that jumped to 3,206. That's not linear growth — it's exponential. This is the case for choosing V4-Pro as the backend for coding agents and developer tools.

That said, a Codeforces rating doesn't mean everything. There's a gap between contest algorithms and real-world code. In production, readability, maintainability, and adherence to team conventions matter. A 3,206 rating doesn't mean you can hand off all code review to the model. Use it as a tool and keep the final judgment human.

Reasoning Benchmarks — Between GPT-5.2 and GPT-5.4

According to DeepSeek's official announcement, V4-Pro scores above GPT-5.2 and below GPT-5.4 on standard reasoning benchmarks covering math, logic, and science. It's one step behind GPT-5.5.

Factor in price, though, and the math shifts. GPT-5.5 output costs $30 per 1M tokens. V4-Pro is $3.48. That's about 8.6x cheaper on output. If you can separate "tasks that absolutely require GPT-5.5" from "tasks where GPT-5.4 level is enough," V4-Pro wins on cost for the latter — decisively.

V4-Flash is one step below V4-Pro on reasoning. DeepSeek claims it approaches GPT-5.2 level. For pipelines where speed and cost are the top priorities, V4-Flash is the realistic choice.

Benchmark numbers are reference points, not final answers. Real application performance varies significantly by domain and prompt design. A model that leads on math reasoning benchmarks may produce different results on tasks that require specialized domain knowledge. Building your own evaluation set and running both models against it is the most accurate approach.

API Pricing — The Difference in Numbers

V4-Pro's API price is $1.74 input and $3.48 output per 1M tokens. GPT-5.5 is $5 input and $30 output. On output tokens, V4-Pro is about 8.6x cheaper. Compared to Claude Opus 4.7 output ($25), the gap is about 7.2x.

V4-Flash is even more extreme. $0.14 input, $0.28 output. That's roughly 107x cheaper than GPT-5.5 on output. Compared to Claude Opus 4.7 output, it's about 89x less. For pipelines handling millions of requests, switching to V4-Flash alone can completely restructure your AI infrastructure cost model.

Running the actual cost simulation makes the gap even clearer. Assume 100,000 requests per day with an average of 500 output tokens each. With V4-Flash, the daily cost is about $14. The same workload on GPT-5.5 is $1,500. That's over $44,000 per month — in a single line item.

Cache hit discounts are also available. DeepSeek's API supports prompt caching, which reduces input costs further when the same system prompt repeats. For RAG pipelines or chatbots with a fixed long system prompt, the actual bill ends up lower than the listed price.

MIT License and Open Weights

The MIT license is one of the most permissive open-source licenses. Commercial use, modification, and redistribution are all allowed. No royalties. Both V4-Pro and V4-Flash were released under these terms.

"Open weights" is different from fully open source. Only the weight files are public. Training data, training code, and the infrastructure used were not released. You're free to host the model yourself or fine-tune it, but reproducing the training process from scratch isn't possible.

The practical implications are significant. You can run it on your own GPU servers without paying API fees. You can fine-tune it for a specific domain to improve performance. You can use it on-premises without sending data outside. In fields like healthcare, legal, and finance where data security is strict, this point is decisive.

There's one thing to verify before applying the MIT license to a business. The license text and DeepSeek's terms of service are separate documents. What the MIT license permits and what DeepSeek's API terms permit may differ. Read both documents yourself before building a commercial service.

Huawei Chip Integration — Working Around Semiconductor Sanctions

DeepSeek V4 was optimized for the Huawei Ascend 910B chip. With the US restricting exports of NVIDIA H100 and H200 to China, DeepSeek built its own kernels and training pipelines for Huawei hardware. They created an infrastructure that runs training and inference without NVIDIA.

This isn't just a technology choice. It's a strategy to build an independent AI infrastructure while working around Western semiconductor sanctions. It also serves as a numerical proof that "you can build frontier-level models without American chips." The Huawei chip hasn't fully caught up to NVIDIA GPU performance yet, but that gap is narrowing fast.

The Huawei Ascend 910B currently performs at roughly 70–80% of the NVIDIA H100 in memory bandwidth and FP16 compute. DeepSeek narrowed the gap with MoE architecture and custom kernels. It's an example of compensating for hardware constraints through software design.

If this strategy succeeds, the implications are significant. Chinese companies with access to Ascend chips could run frontier-level models without NVIDIA. It could mean US semiconductor export restrictions aren't having the intended effect. The growth of the Huawei Ascend ecosystem is becoming a meaningful variable in semiconductor geopolitics.

Calling the DeepSeek API — curl and Python Examples

The DeepSeek API is OpenAI-compatible. The endpoint is different, but the request and response format is identical. In existing OpenAI SDK code, you only need to change base_url and api_key. Migration effort is minimal.

Here's how to call V4-Pro with curl. Change the model value and V4-Flash works the same way. Lower temperature produces more consistent code output.

In Python, use the openai package as-is. Just point base_url at the DeepSeek endpoint. The example below shows a bulk document classification pipeline with V4-Flash. Running the same workload on GPT-5.5 costs about 107x more.

Because the response format is identical to OpenAI, migrating an existing OpenAI-based application to DeepSeek requires minimal code changes. Two environment variables is enough. A gradual migration — DeepSeek in test, GPT-5.5 in production — is also straightforward.

Self-Hosting V4 — Running It on Your Own Server with vLLM

The first question for self-hosting is hardware. Loading V4-Pro (1.6T total parameters) in FP16 requires roughly 3,200GB of VRAM. That's not realistic for most. With 4-bit quantization (GPTQ/AWQ), that drops to about 800GB — roughly 10 H100 80GB cards.

V4-Flash (284B total, 13B active) is far more practical. 4-bit quantization brings it down to about 142GB of VRAM. Two H100 80GB cards are enough. For experimentation, 4–6 RTX 4090 24GB cards are workable. If you're a startup or research lab looking for a realistic starting point for self-hosting, V4-Flash is it.

vLLM is the most widely used serving framework right now. It automatically provides an OpenAI-compatible API. You can switch to self-hosted inference by changing only the endpoint — no code modifications. Here's the basic command to serve V4-Flash with vLLM.

Self-hosting means zero API costs. But GPU server costs and operational overhead kick in. Under 100K requests per day, using the API is probably cheaper. Over 1M requests per day, self-hosting economics improve sharply. If data security requirements exist, self-hosting may be the only option regardless of request volume.

V4-Pro vs V4-Flash — Choosing by Use Case

The situations where V4-Pro is the right call are clear. Complex mathematical reasoning, large-scale code analysis, long document summarization, and precise code generation. If you need Codeforces 3,206-level algorithm implementation, use V4-Pro. It's cheaper than GPT-5.5 and delivers performance close to GPT-5.4.

V4-Flash is for when cost is the top priority. It fits high-throughput classification pipelines, chatbot response generation, short document summarization, and fast prototyping. The 107x cheaper output price versus GPT-5.5 says it all. Both are MIT open weights, so there's no overhead to experimenting with self-hosting either.

There's a difference from a fine-tuning perspective too. V4-Flash has fewer active parameters, so fine-tuning costs and time are lower. If you need a lightweight domain-specialized model, V4-Flash is the rational base. V4-Pro makes sense when you need peak general performance and have the infrastructure to invest in fine-tuning.

Don't pick a model without testing. Build your own evaluation set, run both models against it, and compare the results. Korean language processing quality, domain-specific knowledge, and output format compliance must all be verified through direct testing. Benchmark numbers are a starting point, not a final answer.

Hybrid Pipeline — Combining Three Models

Combining all three models by role lets you optimize for both cost and quality. V4-Flash handles classification, routing, and short summarization. V4-Pro handles complex reasoning, code generation, and long document analysis. GPT-5.5 covers only the small subset of tasks where absolute accuracy is non-negotiable. This three-tier routing alone cuts overall costs significantly.

Keep the routing logic simple to start. Select models based on input token count, keywords, and task type. Rule-based routing is often sufficient without a sophisticated ML router. Add a small classification model to the router later if you need more precision. Here's a basic router in TypeScript.

Run a cost simulation first. Assume 70% of your current GPT-5.5-only pipeline requests are simple classifications. Moving that 70% to V4-Flash cuts total costs substantially. The realistic order is to analyze your actual usage patterns in logs, then design a routing strategy based on what you find.

There's one thing to watch in the hybrid approach. Output format can vary subtly between models. If the next stage of your pipeline expects consistent formatting, add explicit output format instructions to each model's system prompt. Response length and language style also differ between models — verify this in the testing phase.

The Pressure on OpenAI and Anthropic

DeepSeek V4 isn't just a new model release. It's a direct challenge to the premise that "only closed, expensive models can deliver frontier-level performance." An MIT open-weights model with performance near GPT-5.4 at 8.6x cheaper output than GPT-5.5 is now public.

OpenAI and Anthropic have maintained the formula: top performance equals closed model. After V3, V4 narrowed that gap again. At the current pace, open weights reaching parity with closed models within 1–2 years is a credible forecast.

DeepSeek's emergence is forcing price competition on OpenAI and Anthropic. GPT-4o pricing dropped significantly from launch levels over the past year or two. Claude Haiku and GPT-4o mini prices were cut after DeepSeek V3 launched. V4 turns up that pressure another notch.

The practical move is a hybrid approach. Use V4-Flash for everyday tasks, V4-Pro for complex reasoning, and GPT-5.5 only when absolute accuracy is non-negotiable. You can cut API costs significantly while maintaining per-task quality. There's no need to replace everything immediately — but the case for experimenting is solid.

Frequently Asked Questions

Which model should I choose between V4-Pro and V4-Flash?

If you need complex reasoning or long-context analysis, go with V4-Pro. If fast responses matter more — chatbots, summarization, classification pipelines — V4-Flash is the right fit. V4-Flash costs about 12x less per output token than V4-Pro. The most accurate selection criterion is running your own real-world task data through both models, not benchmark numbers alone. Both are MIT open weights, so you can experiment for free.

When do you actually need a 1M token context?

When you need to feed an entire large codebase for analysis. Also when processing hundreds of pages of legal or contract documents in a single pass. Long-running agent systems that need to maintain extended conversation history qualify too. For a standard chat application, 128K is more than enough. The 1M context window isn't something you always use — it removes the ceiling and gives you design flexibility.

Does the MIT license mean unlimited commercial use?

Commercial use, modification, and redistribution are all permitted. No royalties. That said, open weights means only the weight files are public. Training data and training code were not released. The MIT license and DeepSeek's terms of service are separate documents — make sure to read both before building a commercial service.

How impressive is a Codeforces rating of 3,206?

On Codeforces, 2,400 and above is the top Grandmaster tier. 3,206 is above that. It ranks 23rd among all registered human users worldwide — top 0.001% by competitive programming standards. It means the model can solve complex dynamic programming, graph traversal, and mathematical optimization problems at that level. Contest algorithm ability and production code quality don't always map one-to-one, but it's a reliable signal of logical reasoning capability.

Does a hybrid approach using both DeepSeek V4 and GPT-5.5 make sense?

It makes sense from a cost optimization standpoint. Use V4-Flash for everyday summarization and classification, V4-Pro for complex reasoning, and GPT-5.5 only when maximum accuracy is non-negotiable. You cut API costs significantly while maintaining per-task quality. The practical order is to start experimenting on your most expensive pipelines with V4-Flash first, and step up to V4-Pro if quality holds.

Is the DeepSeek API compatible with the existing OpenAI SDK?

Yes. DeepSeek's API uses the same request and response format as OpenAI's API. In the Python openai package, change base_url to https://api.deepseek.com/v1 and api_key to your DeepSeek key — it works immediately. Set the model parameter to deepseek-v4-pro or deepseek-v4-flash as needed. Code changes are minimal.

What hardware do you need to self-host V4-Pro?

Loading V4-Pro (1.6T parameters) in FP16 requires roughly 3,200GB of VRAM. With 4-bit quantization that drops to about 800GB — approximately 10 H100 80GB cards. V4-Flash comes in at about 142GB with 4-bit quantization, which means 2 H100s are sufficient. For research and startup environments, V4-Flash self-hosting is the realistic first step.

What is the difference between DeepSeek V4-Pro and DeepSeek R1?

R1 is a reasoning-specialized model trained with reinforcement learning. It explicitly outputs step-by-step thinking for math and logic problems. V4-Pro is a general-purpose language model covering coding, reasoning, and long-context processing. For pure math and logic reasoning, R1 is stronger. For general-purpose tasks — coding, document processing, API response generation — V4-Pro covers a broader range. The right choice depends on the task.

Closing

DeepSeek V4 proved two things in numbers. Open-weights models can approach frontier-level performance. And AI API costs can keep falling. The Codeforces 3,206 rating and 1M token context show this isn't a "cheap alternative" — it's a performance-validated option.

There's no need to replace your entire stack right now. The realistic order is to test V4-Flash on your most expensive pipelines first, then step up to V4-Pro if quality holds. Both are MIT-licensed, so the overhead to experimenting with self-hosting is zero. The question has shifted from "which model is the best?" to "which model is good enough for this task?"

The figures in this article are based on DeepSeek's official announcement from April 24, 2026. API pricing and model performance are subject to change — always verify on the official site before use.