DeepSeek V4

What it is

DeepSeek V4 is the open-weight release that has done most to push down the price of frontier-grade AI in 2026. DeepSeek released the V4 Preview and open-sourced it on 24 April 2026, in two sizes: DeepSeek V4 Pro, with 1.6 trillion total parameters and 49 billion active at any one time, and V4-Flash, at 284 billion total and 13 billion active. Both carry a one-million-token context window, putting long-document work within reach even at the lower end of the range.

Two later releases grew out of this preview: V4-Flash-0731 entered public beta on 31 July 2026, and the Pro tier reached general availability on 13 August 2026 as DeepSeek V4 Pro, which now has its own page. The GA launch carried a pricing change with it. At 16:00 UTC on 16 August 2026 the flat $0.87 per million output rate ended, and the API moved to peak and off-peak billing: deepseek-v4-pro output now costs $1.98 per million off-peak and $3.96 at peak, and deepseek-v4-flash $0.66 off-peak and $1.32 at peak, with peak set at 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, and weekends billed off-peak throughout (DeepSeek pricing). The Flash line also gained an experimental image-input variant, DeepSeek-V4-Flash-Vision-Exp, released on 21 August 2026 at the same prices as V4-Flash, which DeepSeek’s changelog pitches at agent tasks needing visual understanding. No weights have been published for it.

How the mixture-of-experts design works

The gap between total and active parameters is the key to how DeepSeek V4 keeps costs down. Like other modern large models, V4 is a mixture-of-experts design: only a fraction of its parameters fire for any given token, so V4-Pro runs with 49 billion of its 1.6 trillion parameters active at a time. That is what lets a very large model be served cheaply, because the compute per query tracks the active count rather than the full total.

The two-model split is itself a statement of intent. V4-Pro is the capability play, a very large system meant to compete near the top of the field, while V4-Flash, at 284 billion total and 13 billion active parameters, is the smaller sibling built to run faster and cheaper for high-volume work. Offering both, and open-sourcing them together, lets DeepSeek cover everything from demanding reasoning down to routine, cost-sensitive tasks from a single release. The “Preview” label on V4 also signals that a more polished version is expected to follow. DeepSeek has kept iterating on the smaller sibling in the meantime: a post-training update, V4-Flash-0731, arrived on 31 July 2026 with stronger agent capabilities at the same address and price.

Efficiency engineering

DeepSeek has pushed that efficiency further with a set of new techniques. The model uses a hybrid attention scheme combining what it calls Compressed Sparse Attention and Heavily Compressed Attention, alongside Manifold-Constrained Hyper-Connections. The pay-off shows up at long context: at one million tokens, DeepSeek says V4-Pro needs only 27% of the inference compute and 10% of the memory cache of the previous V3.2. Cutting the cost of long-context inference so sharply is the sort of engineering that makes cheap pricing sustainable rather than a loss-leading stunt.

Capability and price

On capability, DeepSeek positions V4 as open-source state of the art for agentic coding at launch, the multi-step, tool-using work that has become this year’s competitive front line. On world knowledge, the company says V4 trails only Gemini 3.1 Pro across all the models it compared, open or closed. Those are DeepSeek’s own claims and await independent confirmation, but they place DeepSeek V4 unusually close to the closed flagships for a freely available model. Agentic coding is a demanding test in particular, because it asks a model to plan, call tools and correct itself over many steps rather than answer in a single shot, so leading the open-source field there is a claim with real weight behind it.

Price is where the release bites. Output ran at around $0.87 per million tokens from launch, which made DeepSeek V4 the cheapest of the frontier-adjacent set by a wide margin. That flat rate has since gone: from 16 August 2026 the Pro tier bills at $1.98 per million output off-peak and $3.96 at peak, and the Flash tier at $0.66 and $1.32, which still undercuts most frontier APIs even at the peak figure. For developers, that changes the arithmetic of building on a capable model, and it puts pressure on the premium labs to justify charging many times more. Because the weights are open, users can also run the model on their own hardware rather than pay per token at all, an appealing option for anyone with data they would rather not send to a third party.

What to watch

In the wider field, DeepSeek V4 stands as the clearest expression of the open-weight, low-price challenge to the closed US labs. The combination of near-flagship claims, genuine efficiency engineering and pricing far below the incumbents is what makes the release significant. The tests to watch are independent benchmarks and whether the peak and off-peak split spreads to the other cheap open-weight APIs. For the wider picture, see our large language models hub and broader AI coverage.