DeepSeek V4

Overview

DeepSeek V4 is the fourth generation of DeepSeek's open-weight large language model series, released on April 24, 2026. It is not a single model: the DeepSeek API changelog announced two callable models on that date, deepseek-v4-pro and deepseek-v4-flash, both reachable through the OpenAI ChatCompletions interface and the Anthropic Messages interface.

V4-Pro is the frontier tier — 1.6 trillion total parameters with 49 billion activated per token. V4-Flash is the cost tier — 284 billion total with 13 billion activated. Both share the same 1M token context window, the same 384K maximum output, and the same MIT license on Hugging Face.

The generation's defining engineering work is attention. V4 replaces dense attention with a hybrid stack of Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), layered on top of Manifold-Constrained Hyper-Connections (mHC) and trained with the Muon optimizer. DeepSeek reports that at a 1M-token context, V4-Pro "requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2." Long context stops being a premium feature and becomes the default operating mode.

DeepSeek's own model card describes the released checkpoint as "a preview version" of the V4 series, pre-trained on "more than 32T diverse and high-quality tokens."

Capabilities

• Frontier open-weight coding: 80.6 on SWE-bench Verified and a 3206 Codeforces rating put V4-Pro at the top of the open-weight field for software engineering.
• Hybrid thinking: Both models support thinking and non-thinking modes in a single endpoint. Thinking is the default. There is no separate reasoner model to route to.
• Million-token working memory: The hybrid attention stack is what makes a 1M window economically usable rather than a specification-sheet number.
• 384K output: A maximum output length long enough to emit a large refactor, a full technical report, or an extended agent trajectory in one response.
• Dual API surface: Native compatibility with both the OpenAI and Anthropic request formats.
• Self-hostable: MIT-licensed weights on Hugging Face, with vLLM support including an optional speculative decoding module.

Technical Specifications

• Architecture: Mixture-of-Experts Transformer with a Hybrid Attention Architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA)
• Residual structure: Manifold-Constrained Hyper-Connections (mHC)
• Optimizer: Muon
• Pre-training corpus: more than 32T tokens
• Efficiency: at 1M-token context, V4-Pro uses 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2

DeepSeek-V4-Pro-DSpark

DeepSeek-V4-Pro-DSpark appears as a separate Hugging Face repository, but it is not a separate model. DeepSeek states plainly: "DeepSeek-V4-Pro-DSpark is not a new model. It is the same checkpoint with an additional speculative decoding module attached." In vLLM it is enabled with a single flag. Treat it as a serving optimization, not a capability upgrade.

Legacy model IDs

The two legacy API names, deepseek-chat and deepseek-reasoner, are discontinued on July 24, 2026. During the transition window they map to the non-thinking and thinking modes of deepseek-v4-flash respectively. Applications still pointing at those IDs should migrate to explicit V4 model names.

Use Cases

• Repository-scale software engineering: The combination of an 80.6 SWE-bench Verified score, a 1M context window, and a 384K output ceiling suits whole-repo refactors and migrations that need to read broadly before writing.
• Competitive and algorithmic programming: A 3206 Codeforces rating and 93.5 on LiveCodeBench make V4-Pro a strong assistant for algorithm design and contest-style problems.
• Long-document analysis: Legal discovery, regulatory filings, and multi-paper literature review, where a million tokens of context removes most chunking and retrieval plumbing.
• Cost-sensitive high-volume workloads: V4-Flash delivers the same context and output limits at roughly a third of Pro's input price and a third of its output price.
• Private and air-gapped deployment: MIT-licensed weights make on-premises hosting viable for regulated industries.
• Agentic pipelines: Long outputs and hybrid thinking support multi-step tool-using agents without a separate reasoning model.

Performance / Benchmarks

Numbers below are from DeepSeek's own model cards on Hugging Face.

DeepSeek-V4-Pro

DeepSeek-V4-Flash

V4-Flash figures are reported in Think Max mode. Vendor-reported benchmarks are self-selected; treat them as an upper bound and validate on your own evaluation set.

Limitations

• No single "V4" endpoint: You must choose deepseek-v4-pro or deepseek-v4-flash. Code that assumes a generic deepseek-v4 ID will fail.
• Preview checkpoint: DeepSeek describes the released V4-Pro weights as "a preview version" of the series. Behavior may shift in later revisions.
• Legacy ID cutoff: deepseek-chat and deepseek-reasoner stop working on July 24, 2026.
• Serving cost at 1.6T: Even at 49B active parameters, self-hosting V4-Pro requires a substantial multi-GPU cluster. V4-Flash is the practical self-hosting target for most teams.
• Weights, not full openness: MIT covers the weights. Training data and the full training pipeline are not published.
• No published knowledge cutoff: DeepSeek does not state a reliable knowledge cutoff date for V4. Ground time-sensitive queries with retrieval or tool use.
• Benchmarks are vendor-reported: Independent third-party replication of the headline scores is limited.

Pricing & Access

Prices are from DeepSeek's pricing page, quoted per 1M tokens in USD.

The cache-hit price is roughly 1/120th of the cache-miss price on V4-Pro, which makes prompt caching the single highest-leverage optimization for any repeated-context workload. DeepSeek does not publish an off-peak discount for the V4 models.

Access options:

• DeepSeek API — OpenAI ChatCompletions and Anthropic Messages interfaces
• DeepSeek Chat — official web and mobile applications
• Hugging Face — MIT-licensed weights for private deployment

Ecosystem & Tools

• DeepSeek API documentation — reference, changelog, and migration guides
• DeepSeek-V4-Pro on Hugging Face — frontier-tier weights
• DeepSeek-V4-Flash on Hugging Face — efficient-tier weights
• DeepSeek-V4-Pro-DSpark — the same V4-Pro checkpoint with a speculative decoding module for vLLM
• GitHub — inference code and research releases
• vLLM — supported serving path, including DSpark speculative decoding behind a single flag

Community & Resources

• DeepSeek API changelog — the V4-Pro / V4-Flash launch entry and the legacy-ID deprecation notice
• Pricing
• DeepSeek-V4-Pro model card
• DeepSeek on GitHub
• Compare with Kimi K2.6, GLM-5.2, Qwen3.7-Max, and Ling-2.6-1T