DeepSeek Releases V4-Pro with 1.6T Parameters, 1M Token Context at 27% Inference Cost of V3
DeepSeek has released two Mixture-of-Experts models: V4-Pro with 1.6 trillion parameters (49B activated) and V4-Flash with 284B parameters (13B activated), both supporting 1 million token context windows. V4-Pro requires only 27% of inference FLOPs and 10% of KV cache compared to V3.2 at 1M token context, trained on over 32 trillion tokens.
DeepSeek V4 Pro at a glance
How DeepSeek V4 Pro ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| AA Intelligence Index | 23.1 index | 36 of 49 | Artificial Analysis, Nov 2026 ↗ |
| Arena Elo | 1464 Elo | 41 of 141 | LMArena, Nov 2026 ↗ |
| GPQA | 92.8% | 20 of 163 | Artificial Analysis, Oct 2026 ↗ |
| SWE-bench Verified | 77.4% | 28 of 80 | Vals AI, Sep 2026 ↗ |
| Speed (tok/s) | 95 tok/s | 48 of 101 | Artificial Analysis, Nov 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
DeepSeek Releases V4-Pro with 1.6T Parameters, 1M Token Context at 27% Inference Cost of V3
DeepSeek has released two new Mixture-of-Experts language models: DeepSeek-V4-Pro with 1.6 trillion parameters (49 billion activated) and DeepSeek-V4-Flash with 284 billion parameters (13 billion activated). Both models support a context length of 1 million tokens.
Efficiency Gains Through Hybrid Attention
The V4 series introduces a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). In 1M-token context settings, V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared to DeepSeek-V3.2.
Both models were pre-trained on more than 32 trillion diverse tokens, followed by a two-stage post-training pipeline: independent domain-specific expert cultivation through supervised fine-tuning and reinforcement learning with GRPO, then unified model consolidation via on-policy distillation.
Benchmark Performance
DeepSeek-V4-Pro-Base achieves 90.1% on MMLU (5-shot), 73.5% on MMLU-Pro, 76.8% on HumanEval (0-shot), and 51.5% on LongBench-V2. The model scores 55.2% on Simple-QA verified and 62.6% on FACTS Parametric, indicating significant knowledge capability improvements.
DeepSeek-V4-Pro-Max, the maximum reasoning effort mode, achieves a 3206 Codeforces rating, 93.5% on LiveCodeBench, 89.8% on IMOAnswerBench, and 90.2% on Apex Shortlist. On agentic tasks, it scores 80.6% on SWE Verified, 67.9% on Terminal Bench 2.0, and 83.4% on BrowseComp.
Three Reasoning Modes
Both models support three reasoning effort modes:
- Non-think: Fast, intuitive responses for routine tasks
- Think: Conscious logical analysis with visible reasoning chains
- Think Max: Maximum reasoning capability with extended thinking budget
The Flash-Max variant achieves comparable reasoning performance to Pro when given larger thinking budgets, though it trails on pure knowledge tasks and complex agentic workflows due to its smaller parameter scale.
Technical Architecture
V4 series incorporates Manifold-Constrained Hyper-Connections (mHC) to strengthen residual connections, enhancing signal propagation stability across layers. The models use the Muon optimizer for faster convergence and greater training stability.
Models are available in FP8 mixed precision (base versions) and FP4 + FP8 mixed precision (post-trained versions), where MoE expert parameters use FP4 and other parameters use FP8.
Availability
All four model variants (V4-Pro-Base, V4-Pro, V4-Flash-Base, V4-Flash) are available on HuggingFace and ModelScope. Pricing has not been disclosed.
Note: The DeepSeek-V4-Pro-DSpark checkpoint is not a new model but the same V4-Pro checkpoint with an additional speculative decoding module for inference optimization.
What This Means
DeepSeek's 73% reduction in inference costs at 1M context length addresses a critical bottleneck in long-context applications. The V4-Pro-Max performance on coding benchmarks (3206 Codeforces rating) and math reasoning (89.8% IMOAnswerBench) positions it competitively with frontier closed-source models like Claude Opus 4.6 and GPT-5.4. The three-tier reasoning mode system provides practical flexibility for balancing speed and accuracy based on task complexity. The open-source release of models this large (1.6T parameters) with competitive performance represents a significant shift in accessibility to frontier-level capabilities.
Related Articles
Runway releases Praxis-1 robot model with open weights under Apache 2.0, tops World Action Models on RoboLab-120
Runway has released Praxis-1, a robot control model, with open weights and source code under the Apache 2.0 license. It scores 46.2 percent on Nvidia's RoboLab-120 leaderboard, ahead of FLUX 3 Action at 42.9 percent, in tests Runway ran itself.
DeepSeek opens Search, Fetch, OCR, ASR and TTS as standalone APIs, with tool fees billed on top of tokens
DeepSeek has made the capabilities behind its app available as standalone APIs: Search and Fetch, OCR, ASR and TTS. It also open-sourced chat-demo, a static web client that connects to the API from the browser. Tool use is billed on top of tokens.
Translated releases Lara Mini, a 7.2B-parameter MoE research model with 600M active parameters
Translated announced Lara Mini on November 12, 2026, a research model trained from scratch on 6 trillion tokens across English and 44 other languages, including all 24 official EU languages. The mixture-of-experts system has 7.2 billion parameters and activates 600 million per token.
Analysis finds Blackwell fingerprints in six open-weight models from four Chinese labs
Unemployment Labs reports that six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry numerical patterns matching NVIDIA Blackwell GPUs. Earlier models from the same labs match the older Hopper generation. The analysis says it does not establish where the chips were located or whether export controls were violated.
Comments
Loading...