A year after its “Sputnik moment,” DeepSeek just released something harder to ignore than a benchmark headline.
On April 24, 2026 — exactly one year after DeepSeek’s V3/R1 models shook global AI markets — the Chinese lab released preview versions of two new open-weight models: V4-Pro and V4-Flash. The timing is deliberate. The message is clear. And for developers evaluating their AI stack in 2026, the implications deserve attention.
What Hit the Market
Both models are Mixture-of-Experts architectures. V4-Pro loads 1.6 trillion total parameters with 49 billion active per token — making it, by parameter count, the largest open-weight model ever released, surpassing Kimi K2.6 (1.1B) and GLM-5.1 (754B). V4-Flash is the lighter sibling: 284 billion total / 13 billion active.
Both are released under MIT license. Both support a native context window of 1 million tokens. Both are available on Hugging Face, DeepSeek’s API, and chat.deepseek.com (Expert Mode = V4-Pro, Instant Mode = V4-Flash).
The community started forking and quantizing within hours of release.
The Architecture Story: Efficiency Over Brute Force
The real news isn’t the parameters — it’s what DeepSeek did with attention.
V4 introduces a Hybrid Attention Architecture that combines two mechanisms: Compressed Sparse Attention (CSA), which maintains a compact KV cache plus a top-k sparse selector, and Heavily Compressed Attention (HCA), which condenses many tokens into a single entry. The alternation between both is what makes the 1M token window operationally viable and not just theoretically available.
The numbers are concrete: with 1M token context, V4-Pro requires only 27% of the inference FLOPs of DeepSeek V3.2 and just 10% of its KV cache. V4-Flash goes even further: 10% of FLOPs and 7% of KV cache.
For developers, this translates directly. Loading an entire large repository as a single prompt — something that with previous models was expensive and frequently incoherent — becomes a realistic workflow. Coherence improvements across long context sessions aren’t marketing; they’re an architectural consequence.
V4 also introduces Manifold-Constrained Hyper-Connections (mHC), a technique that strengthens residual connections to improve signal propagation stability across the model’s multiple layers without sacrificing expressivity.
Where Benchmarks Land
V4-Pro-Max (maximum effort reasoning mode) scores 93.5 on LiveCodeBench — a new high for open-weight models — and a Codeforces rating of approximately 3,206, which DeepSeek places around rank 23 among human competitors in competitions. On formal reasoning Putnam-2025, it achieves a perfect 120/120.
Honest positioning: V4-Pro-Max exceeds GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. It falls marginally below GPT-5.4 and Gemini-3.1-Pro — according to DeepSeek’s own characterization, the model is approximately 3 to 6 months behind the latest frontier models in terms of development.
It’s worth naming that gap. And also what it costs to close it.
The Pricing Argument Is the Real Story
This is where the strategic picture becomes impossible to ignore.
V4-Flash: USD 0.14/million input tokens, USD 0.28/million output. V4-Pro: USD 1.74/million input, USD 3.48/million output.
Simon Willison, whose model pricing comparisons are among the most cited in the industry, notes that V4-Flash is the cheapest model in the current small model tier — even below OpenAI’s GPT-5.4 Nano. V4-Pro is the cheapest among larger frontier-class models, with significant margins over OpenAI and Anthropic equivalents.
DeepSeek attributes this to genuine architectural efficiency gains, not subsidized pricing. Though it’s worth noting the company acknowledged throughput constraints at launch due to “high-end compute limitations,” and indicated prices could drop further when 950 new Huawei Ascend supernodes come online later in 2026.
This isn’t a model competing to top every leaderboard. It’s a model with pricing such that the cost of not evaluating it is significant.
What This Means If You’re Building
For API consumers: The V4-Flash tier is the first worth benchmarking. At USD 0.14/M input, it’s accessible for high-volume applications where cost has been the limiting factor for using frontier-class models. For complex agentic workflows requiring deep reasoning, V4-Pro-Max is the natural comparison point against closed-source equivalents.
For long-context workflows: The 1M token window isn’t new in 2026 — but the efficiency story is. Loading a large repository, a complete documentation set, or an extended multi-session conversation without coherence degradation or prohibitive inference costs is now a practical consideration, not theoretical.
For those wanting to host locally: The reality check matters. V4-Pro, at approximately 865 GB on disk in mixed precision FP4/FP8, requires at minimum eight 80 GB H100 GPUs with NVLink for a realistic deployment — this is data center territory. V4-Flash at 160 GB is another story: a quantized version could run on a 128 GB MacBook Pro M5, and the Unsloth team started releasing quantized variants within hours of launch.
One dimension that doesn’t disappear: Data sovereignty. Prompts sent to DeepSeek’s hosted API traverse Chinese infrastructure. For teams where data residency matters, the alternatives are: host locally (total sovereignty, real hardware cost) or use a third-party API provider like OpenRouter, which serves the same weights from U.S. or European infrastructure.
The Structural Question Worth Asking
DeepSeek’s real achievement isn’t defeating GPT-5.4 on a leaderboard. It’s demonstrating, for the second year running, that frontier AI cost curves can break faster than incumbent pricing models assume.
A model that performs at 85–90% of frontier capability at 10–15% of the cost doesn’t need to win the benchmarking war to reshape procurement decisions, enterprise architecture discussions, and the competitive dynamics of all AI-native products built on top of these APIs.
Open-weight status accelerates all of this. The forks, quantizations, and fine-tunes from the community that will appear over the coming weeks will extend the practical reach of these models far beyond what DeepSeek publishes today.
Whether you use V4 or not, the models your competitors build on top of it will be worth following closely.
DeepSeek-V4-Pro and V4-Flash are available as open weights on Hugging Face under the MIT license and through DeepSeek’s API and chat.deepseek.com.