DeepSeek V4
Last updated: April 26, 2026
Specs
| Vendor | DeepSeek |
|---|
| Released | April 24, 2026 |
|---|
| Context window | 1M tokens |
|---|
| Input | $1.74 / M tokens |
|---|
| Output | $3.48 / M tokens |
|---|
Benchmarks
| mmlu | 91.5 |
|---|
| gpqa | 73 |
|---|
| humanEval | 91 |
|---|
| arenaElo | 1395 |
|---|
| sweBench | 62 |
|---|
| sweBenchPro | 54 |
|---|
| liveCodeBench | 73 |
|---|
| aiderPolyglot | 70 |
|---|
| terminalBench | 40 |
|---|
| mmluPro | 85 |
|---|
| hle | 18 |
|---|
| aime | 90 |
|---|
| math500 | 97 |
|---|
| mmmu | 76 |
|---|
Strengths
- 1.6T-parameter MoE flagship under MIT license: fully open-source
- 1M-token context with hybrid sparse attention (~10% of V3.2 KV cache)
- Roughly 1/6th the cost of GPT-5.5 and Claude Opus 4.7 for similar frontier-tier output
Weaknesses
- Lags top frontier models on Chatbot Arena (creative writing, instruction following)
- No native audio modality
- Tool-use ecosystem less mature than OpenAI / Anthropic
Best for
- STEM, math and graduate-level reasoning at scale
- Cost-sensitive 1M-context workloads (document + codebase analysis)
- Self-hosted enterprise deployments with data residency requirements
Compare with other AI models