Model Details
DeepSeek-V4 Pro Max
Quick answer: DeepSeek-V4 Pro Max is DeepSeek's February 2026 most powerful model, scoring 93.5% on LiveCodeBench, 87.5% on MMLU-Pro, and 48.2% on Humanity's Last Exam. A proprietary (non-open-weight) model, it represents DeepSeek's shift to closed weights for their top-tier offering.
At a Glance
Where DeepSeek-V4 Pro Max leads
- 93.5% LiveCodeBench — near-ceiling competitive coding performance
- 87.5% MMLU-Pro — state-of-the-art academic knowledge
- 48.2% Humanity's Last Exam — strong frontier task performance
- Strongest DeepSeek model at its release date
- Low-cost API access (DeepSeek's characteristic pricing)
Where it lags
- Proprietary: closed weights, no self-hosting
- No MIT/Apache open license (unlike V3 and V3.2)
- Text-only
Best for: Maximum DeepSeek capability for coding and academic tasks; production API deployments requiring top-tier performance.
What DeepSeek-V4 Pro Max Is
DeepSeek-V4 Pro Max (released February 2026) is DeepSeek's flagship model in the V4 generation. Unlike prior DeepSeek models (V2, V3, R1) which were open-weight with MIT or permissive licenses, V4 Pro Max is proprietary — accessible only via the DeepSeek API.
The model achieves 93.5% LiveCodeBench — among the highest coding scores publicly reported — and 87.5% MMLU-Pro, placing it at the frontier of AI capabilities at its release date. The 48.2% Humanity's Last Exam score is also notable on this challenging benchmark.
For teams seeking open-weight DeepSeek models, DeepSeek-V3 (MIT) or DeepSeek-V3.2 Experimental remain available. V4 Pro Max is for teams prioritising maximum performance over licensing flexibility.
Specifications
| Field | Value |
|---|---|
| Organization | DeepSeek |
| License | Proprietary (API only) |
| Release date | February 2026 |
| Knowledge cutoff | September 2025 |
| Modality | Text only |
Pricing
Available via DeepSeek API. Refer to the DeepSeek pricing page for current rates.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| LiveCodeBench | 93.5% | Benchgen evaluation | 2026-02 |
| MMLU-Pro | 87.5% | Benchgen evaluation | 2026-02 |
| Humanity's Last Exam | 48.2% | Benchgen evaluation | 2026-02 |
DeepSeek-V4 Pro Max vs Alternatives
| Model | LiveCodeBench | MMLU-Pro | License |
|---|---|---|---|
| DeepSeek-V4 Pro Max | 93.5% | 87.5% | Proprietary |
| DeepSeek-V4 Flash Max | 91.6% | 86.2% | Proprietary |
| DeepSeek-V3.2 Experimental | 74.1% | 85% | MIT |
| DeepSeek-V3 | 27.2% | — | MIT |
V4 Pro Max vs V4 Flash Max: higher LiveCodeBench (93.5% vs 91.6%) and MMLU-Pro (87.5% vs 86.2%) at higher cost. For lower-cost frontier-quality coding, Flash Max. For open weights, V3.2 Exp (MIT, 74.1% LiveCodeBench).
Related Models & Benchmarks
- Compare: DeepSeek-V4 Flash Max · DeepSeek-V3.2 Experimental · DeepSeek-V3
- Benchmarks: LiveCodeBench · MMLU-Pro · Humanity's Last Exam
Frequently Asked Questions
What is DeepSeek-V4 Pro Max?
DeepSeek's February 2026 proprietary flagship scoring 93.5% LiveCodeBench, 87.5% MMLU-Pro, and 48.2% HLE. DeepSeek's most capable model with closed weights.Is DeepSeek-V4 Pro Max open source?
No — V4 Pro Max is proprietary, accessible only via the DeepSeek API. For open-weight DeepSeek models, use V3 (MIT) or V3.2 Experimental (MIT).Specs from DeepSeek's V4 release (February 2026) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.