DeepSeek V4: A 1.6T Open-Weight MoE with 1M Context

DeepSeek released V4 on April 24, 2026, and unlike the quiet R1 point updates, this one is a generational jump. It comes in two sizes: V4-Pro, a 1.6-trillion-parameter mixture-of-experts model, and V4-Flash at 284 billion. Both are open-weight under an MIT license with a 1-million-token context window by default. Frontier-adjacent benchmarks plus weights you can actually download is the rare part.

Two Models, One Architecture

Both V4 models are sparse mixture-of-experts, so the total parameter count overstates what actually runs per token. V4-Pro activates roughly 49B of its 1.6T parameters on any given token; V4-Flash activates about 13B of 284B. That’s how DeepSeek keeps inference costs down while still claiming frontier-class quality: you pay compute for the active path, not the full model.

In the API you select them with deepseek-v4-pro and deepseek-v4-flash, and DeepSeek kept the endpoint compatible with both the OpenAI ChatCompletions and Anthropic message formats, so dropping V4 into an existing integration is usually a base-URL-and-model-name change.

The Benchmarks

The headline number is SWE-bench Verified, where DeepSeek reports 80.6% for V4-Pro. That lands roughly 0.2 of a percentage point behind Claude Opus 4.7 (80.8%), close enough that the practical difference on real software-engineering tasks is negligible, and notable mainly because it’s an open-weight model trading blows with proprietary frontier releases. On knowledge and reasoning, V4-Pro posts 87.5% on MMLU-Pro, 90.1% on GPQA Diamond, and 92.6% on GSM8K.

As always with vendor-reported figures, treat them as a starting point rather than gospel and test on your own workload. But the pattern is consistent with what third-party trackers have measured: V4 is genuinely in the frontier conversation, not a budget alternative that benchmarks well and disappoints in production.

Why the Long Context Is Actually Usable

A 1M-token window is only useful if it’s affordable to fill. V4’s main architectural story is a hybrid attention scheme: DeepSeek pairs a compressed sparse attention path with a heavily compressed one to cut the cost of long-context inference. In the 1M-token setting, the company says V4-Pro needs only about 27% of the per-token inference FLOPs and 10% of the KV cache compared with V3.2. That’s the difference between a context window you can use on every request and one you reach for only when you have to.

The Economics

This is where V4 gets genuinely disruptive for anyone weighing build versus buy. The weights are downloadable, so if you have the hardware you can serve it yourself with no per-token cost and no data leaving your infrastructure. And the hosted API is priced aggressively: on May 22, 2026, DeepSeek made its 75% V4-Pro price cut permanent, bringing standing rates to $0.435 per million input tokens, $0.87 per million output, and roughly $0.0036 per million on cache hits, about a quarter of the original list price. DeepSeek framed it not as a promotion but as an efficiency gain passed through to developers. That’s a fraction of what the proprietary frontier APIs charge for comparable quality, and for high-volume agentic workloads, where you’re firing off a lot of reasoning calls and the absolute ceiling isn’t required, the gap compounds fast.

What It Means

V4 is the clearest sign yet that the open-weight frontier hasn’t fallen far behind the closed one. A model you can download, that posts SWE-bench numbers within a rounding error of Claude Opus, at a fraction of the API cost, changes the build-versus-buy math for a lot of teams, and it sits near the top of our best open source models for coding roundup, especially anyone with data-residency constraints that make sending requests to a US cloud a non-starter.

If you’ve been defaulting to a proprietary API for reasoning-heavy work, V4 is worth a serious evaluation. And if you’re already self-hosting DeepSeek derivatives, the V4 bases are the ones to build on next.