DeepSeek V4 Pro and V4 Flash

2026-04-24 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases V4-Pro (1.6 trillion parameters, 49 billion active) and V4-Flash (284 billion, 13 billion active) in preview, both with a one-million-token context, under the MIT license.

### Context V3.2 (December 2025) had brought the V3 line to the announced level of the closed models of the time. V4 changes scale and architecture. As of April 24, 2026, the closed models compared by the vendor are Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro. The goal is a one-million-token context at a sustainable compute cost, and an open model able to act as a coding agent. ### What the model brings - A hybrid attention combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA): at 1 million tokens, V4-Pro requires only 27% of the per-token FLOPs and 10% of the KV cache of V3.2, according to the model card. - Reinforced residual connections (Manifold-Constrained Hyper-Connections, mHC) to stabilize signal propagation between layers. - The Muon optimizer, chosen for faster convergence and more stable training. - A two-stage post-training: domain experts (SFT and RL with GRPO), then merging by on-policy distillation. - Three reasoning modes (no thinking, high thinking, maximum thinking) on both models. - A 1-million-token context now the default on all official services. ### Architecture and specifications | | V4-Pro | V4-Flash | |---|---|---| | Total parameters | 1.6 T | 284 B | | Active parameters | 49 B | 13 B | | Pre-training tokens | 33 T (post) | 32 T (post) | | Context | 1 M | 1 M | | Precision | FP4 (MoE experts) + FP8 | FP4 (MoE experts) + FP8 | - Modalities: text. Number of experts and cutoff: not disclosed in the sources reviewed. - License: MIT (code and weights), commercial use allowed. Base versions (Base) also published. - Weights: collection huggingface.co/deepseek-ai/deepseek-v4. Technical report: arXiv 2606.19348. - Settings: temperature 1.0, top_p 1.0; a context of at least 384K is recommended for maximum mode. ### Cost and access | API (USD per million tokens) | Input, cache hit | Input, cache miss | Output | |---|---|---|---| | deepseek-v4-pro | 0.145 | 1.74 | 3.48 | | deepseek-v4-flash | 0.028 | 0.14 | 0.28 | Launch pricing, replaced on August 16, 2026 by a schedule with peak and off-peak rates (see the V4-Pro-0813 entry). The old identifiers `deepseek-chat` and `deepseek-reasoner` were due to be retired three months later (July 24, 2026). Self-hosting hardware: not disclosed apart from a recommended context of 384K. ### Benchmarks | Benchmark | Opus-4.6 Max | GPT-5.4 xHigh | Gemini-3.1-Pro High | V4-Pro Max | |---|---|---|---|---| | MMLU-Pro | 89.1 | 87.5 | 91.0 | 87.5 | | GPQA Diamond | 91.3 | 93.0 | 94.3 | 90.1 | | Humanity's Last Exam | 40.0 | 39.8 | 44.4 | 37.7 | | LiveCodeBench | 88.8 | not published | 91.7 | 93.5 | | Codeforces (rating) | not published | 3168 | 3052 | 3206 | | Terminal Bench 2.0 | 65.4 | 75.1 | 68.5 | 67.9 | | SWE Verified | 80.8 | not published | 80.6 | 80.6 | | SWE Pro | 57.3 | 57.7 | 54.2 | 55.4 | | BrowseComp | 83.7 | 82.7 | 85.9 | 83.4 | | MRCR 1M | 92.9 | not published | 76.3 | 83.5 | Source: DeepSeek-V4-Pro model card on Hugging Face, scores measured and published by the vendor. ### Versus closed models At the time of its release, the vendor compares V4-Pro Max to Claude Opus 4.6 Max, GPT-5.4 xHigh and Gemini 3.1 Pro High. V4-Pro is ahead on LiveCodeBench (93.5 versus 88.8 and 91.7) and Codeforces (3206 versus 3168 and 3052). It is on par with Gemini 3.1 Pro on SWE Verified (80.6 on both sides) and close to Opus 4.6 (80.8). It remains behind on knowledge (MMLU-Pro 87.5 versus 91.0 for Gemini), on GPQA Diamond (90.1 versus 94.3) and on long context (MRCR 1M at 83.5 versus 92.9 for Opus 4.6). The post states that V4-Pro trails only Gemini 3.1 Pro in world knowledge among the compared models. On price, output is billed at 3.48 dollars per million tokens at launch; as the closed models' prices do not appear in the sources, the gap is not quantified here. ### Key takeaways V4 moves open source into the era of a million tokens by default and to coding-agent scores close to the best closed models of the moment. V4-Flash, about twelve times cheaper on output than V4-Pro at launch (0.28 versus 3.48 dollars), serves as the everyday model.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Contexte 1M, Agents