DeepSeek V4

DeepSeek V4 drops: 1M context, MIT weights, and a trillion-parameter question

On April 24, 2026, DeepSeek released the V4 preview family — two Mixture-of-Experts models, both open-weighted on Hugging Face under the plain MIT license, both with a 1M-token context window and dual Thinking / Non-Thinking modes. DeepSeek-V4-Pro is a 1.6-trillion parameter model with 49B active per token; DeepSeek-V4-Flash is 284B total with 13B active. The release notes and weights went live the same day.

This is the direct successor to DeepSeek V3, the December 2024 model that made “frontier performance for under $6M” a conversation. V4 continues the same research line — DeepSeekMoE routing plus Multi-head Latent Attention — scaled up 2.4x on total parameters and 8x on context length, and it moves the license from the bespoke DeepSeek License to straight MIT. That last change matters more than it looks: MIT means quantizers, distillers, and fine-tuners can build on these weights without reading a custom agreement first.

What’s in the weights

• V4-Pro: 1.6T total parameters, 49B active per token — DeepSeek’s claim is performance rivaling the top closed models
• V4-Flash: 284B total, 13B active — the efficiency play, targeting the tier V3 (671B/37B active) used to occupy with a smaller active footprint
• 1M-token context on both, up from V3’s 128K — the headline is “the era of cost-effective 1M context”
• Dual modes: Thinking (reasoning traces, R1-style) and Non-Thinking in one set of weights, so there is no separate V4 vs R-series split at this tier
• MIT license, full weights on Hugging Face, API model IDs deepseek-v4-pro and deepseek-v4-flash

Hosted API pricing at launch sat well below the closed frontier — roughly $1.74 per million input tokens and $3.48 per million output tokens for V4-Pro as of July 2026, per DeepSeek’s published rates. Cheap hosted inference is not sovereignty, but it does keep pressure on everyone else’s prices, and that pressure is a pleb dividend even if you never touch their API.

Can a pleb actually run it?

Honest math, house rule: a Q4 quant needs roughly 0.6 GB per billion total parameters, plus context overhead — and a 1M-token KV cache is its own memory monster even with MLA compression.

• V4-Pro (1.6T): ~960 GB at Q4. That is not a rig, that is a rack. Multi-node territory — even eight 96 GB cards don’t get you there with context. For home purposes, V4-Pro is an API model that happens to have downloadable weights. The weights being open still matters (auditability, distillations, no rug-pulls), but nobody is running this on a 120V circuit.
• V4-Flash (284B): ~170 GB at Q4. This is the interesting one. A 192 GB Mac Studio M3 Ultra can hold it — but Apple withdrew the 256/512 GB tiers and the 192 GB config is used-market only now, so check the M3 Ultra page before planning around one. A pair of 128 GB unified-memory boxes (Strix Halo, DGX Spark) linked together also lands in range. Only 13B active per token means that once the weights fit, tokens come out at a usable pace even on bandwidth-limited hardware.

For most home rigs, the practical sovereign stack is still built from the 20–120B class — see the best local LLMs for 2026 rundown. V4’s local story will really start when the community distillations land, the same way R1’s distills brought reasoning to 24 GB cards.

What it means for the sovereign stack

Three takeaways. First, the open-weight frontier did not close — the same spring Meta moved its frontier effort behind an API, DeepSeek shipped a trillion-parameter model under MIT. Whose roadmap you depend on matters. Second, 1M context under an open license changes what local RAG pipelines have to do: a whole codebase or a year of records fits in the window, no retrieval gymnastics. Third, V4-Flash’s 284B/13B-active shape is a preview of where pleb-relevant models are heading — huge sparse capacity, small active compute, sized for unified-memory boxes instead of GPU stacks.

Where it sits against the rest of the open field — licenses, VRAM classes, and jurisdiction — is mapped in the open-weight comparison table. Credit where due: V4 stands on Transformer, Switch Transformer, Mixtral’s open-MoE precedent, and DeepSeek’s own V2/V3/R1 line — this is compounding public research, which is exactly why open weights are worth defending.