DeepSeek V4 Pro is the flagship of DeepSeek's V4 family, released on April 24, 2026, and the strongest argument yet that open-weight models compete at the frontier. It is a 1.6-trillion-parameter mixture-of-experts model that activates just 49B parameters per token, pairs a 1M-token context window with up to 384K tokens of output, and ships under a plain MIT license — no usage thresholds, no acceptable-use gate.
The headline engineering story is the attention stack. V4 combines Compressed Sparse Attention (CSA), which compresses KV entries 4× along the sequence with softmax-gated pooling, with Heavily Compressed Attention (HCA) at 128× compression. The practical result: at 1M-token context, V4 Pro needs roughly 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2. Long context stopped being a luxury and became the default across DeepSeek's services.
The API speaks both the OpenAI ChatCompletions and Anthropic formats, so it drops into tools like Claude Code and OpenCode without a proxy — a deliberate play for developers already inside proprietary ecosystems.
DeepSeek V4 specifications
| Vendor | DeepSeek (China) |
|---|---|
| Architecture | Mixture-of-Experts with hybrid sparse attention (CSA + HCA) |
| Total parameters | 1.6T |
| Active parameters | 49B |
| Context window | 1M tokens |
| Max output | 384K tokens |
| Modalities | text |
| License | MIT — weights free to download, self-host, fine-tune, and use commercially |
| Release date | 2026-04-24 |
| Weights | Hugging Face · GitHub |
DeepSeek V4 API pricing
| Direction | Price per 1M tokens |
|---|---|
| Input | $1.74 |
| Output | $3.48 |
Standard first-party API rate. Periodic promotional windows drop it ~75% (to $0.435/$0.87). Cache hits are billed at one tenth of standard input. Verified July 2026.
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| Inference FLOPs vs V3.2 (1M-token context) | 27% | single-token inference cost relative to DeepSeek-V3.2 |
| KV cache vs V3.2 (1M-token context) | 10% |
Running DeepSeek V4 locally
Not a consumer-hardware model. At 1.6T total parameters, self-hosting V4 Pro means a multi-GPU H100/H200-class node (or larger) even with aggressive quantization. Most teams run it via the first-party API or a hosted inference provider; if you want DeepSeek locally, start with V4 Flash or a distilled R1 variant.
What DeepSeek V4 is best for
- Frontier-level reasoning and coding on an open license
- Very long documents and repositories (1M-token context as the default)
- Teams that need Anthropic/OpenAI API compatibility with open-weight economics
Frequently asked questions
Is DeepSeek V4 open source?
DeepSeek V4 is open-weight under the MIT license: the weights are free to download, self-host, fine-tune, and use commercially with no revenue or user thresholds. The full training data and pipeline are not published, which is why 'open-weight' is the more precise term than 'open source'.
What is the difference between DeepSeek V4 Pro and V4 Flash?
V4 Pro is the 1.6T-parameter flagship (49B active per token); V4 Flash is a 284B-parameter model with 13B active, priced at $0.14/$0.28 per million tokens. Both default to the 1M-token context window. Flash trades peak quality for roughly 12× lower cost.
How much does the DeepSeek V4 API cost?
As of July 2026, V4 Pro costs $1.74 per million input tokens and $3.48 per million output tokens on the first-party API, with periodic promotional discounts of ~75% and cache hits billed at one tenth of input. V4 Flash costs $0.14/$0.28.
Can I run DeepSeek V4 locally?
V4 Pro's 1.6T total parameters put it beyond consumer hardware — plan on a multi-GPU H100/H200-class server even quantized. For local use, V4 Flash (284B total / 13B active) quantized, or the older distilled R1 models, are the realistic options.