DeepSeekopen-weight
Released: 2026-04-23DeepSeek V4-Flash
DeepSeek high-throughput low-latency MoE model delivering industry-record price-to-performance at $0.14/1M tokens.
Intelligence Index
88.0/ 100
Calibrated multi-domain compositeThroughput Speed
190tok/sec
Output streaming throughputTime To First Token
230ms
Initial chunk server latencyLiveBench Contam-Free
74.5%
September 2026 suite scoreLiveBench Multi-Domain Evaluation Breakdown
Monthly refreshed contamination-free test setsLogical Reasoning84.0%
Coding & Repo Repair75.5%
Mathematics (AIME)86.0%
Data Analysis & Tables72.0%
Language & Comprehension68.0%
Instruction Following62.0%
Sourced Benchmark Evaluations (61)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| ARC-AGI-2 | Nearing Saturation | 67.5% | 75 | 2026-09-12 | independent | Source ↗ |
| ARC-AGI-3 | Active | 56.2% | 93 | 2026-09-12 | independent | Source ↗ |
| FrontierMath | Active | 24.1% | 98 | 2026-09-10 | independent | Source ↗ |
| PutnamBench | Active | 44.7% | 90 | 2026-09-09 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 51.3% | 20 | 2026-09-07 | independent | Source ↗ |
| Omni-MATH | Deprecated | 63.6% | 10 | 2026-09-07 | independent | Source ↗ |
| Soohak | Active | 65.4% | 98 | 2026-09-05 | independent | Source ↗ |
| FrontierCode | Active | 43.8% | 98 | 2026-09-05 | independent | Source ↗ |
| Humanity's Last Exam | Active | 38.1% | 93 | 2026-09-03 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 52.7% | 75 | 2026-09-03 | independent | Source ↗ |
| AIME | Nearing Saturation | 81.7% | 75 | 2026-09-03 | independent | Source ↗ |
| OlympiadBench | Active | 70.5% | 80 | 2026-09-02 | independent | Source ↗ |
| MathArena | Active | 68.8% | 99 | 2026-09-01 | independent | Source ↗ |
| FireBench | Active | 52.1% | 95 | 2026-09-01 | independent | Source ↗ |
| Terminal-Bench | Active | 52.9% | 80 | 2026-09-01 | independent | Source ↗ |
| OpenAI-MRCR | Active | 73.9% | 80 | 2026-08-31 | independent | Source ↗ |
| SimpleQA | Active | 45.8% | 90 | 2026-08-31 | independent | Source ↗ |
| CodeContests | Deprecated | 49.1% | 10 | 2026-08-31 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 50.8% | 80 | 2026-08-30 | independent | Source ↗ |
| SWE-bench Pro | Active | 46.8% | 90 | 2026-08-30 | independent | Source ↗ |
| LongBench-Pro | Active | 72.2% | 80 | 2026-08-30 | independent | Source ↗ |
| ZeroSCROLLS | Nearing Saturation | 70.4% | 78 | 2026-08-29 | independent | Source ↗ |
| MultiPL-E | Deprecated | 69.5% | 10 | 2026-08-29 | independent | Source ↗ |
| LiveCodeBench | Saturated | 61.9% | 39 | 2026-08-28 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 42.2% | 80 | 2026-08-28 | independent | Source ↗ |
| MATH | Saturated | 84.3% | 25 | 2026-08-28 | independent | Source ↗ |
| InfiniteBench | Nearing Saturation | 74.8% | 65 | 2026-08-28 | independent | Source ↗ |
| Mercury | Deprecated | 66.4% | 10 | 2026-08-27 | independent | Source ↗ |
| SCROLLS | Nearing Saturation | 70.4% | 65 | 2026-08-27 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 81% | 75 | 2026-08-27 | independent | Source ↗ |
| RULER | Nearing Saturation | 81% | 84 | 2026-08-26 | independent | Source ↗ |
| TabMWP | Saturated | 79.1% | 25 | 2026-08-26 | independent | Source ↗ |
| CyberSecEval | Deprecated | 68% | 10 | 2026-08-26 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 58.9% | 25 | 2026-08-25 | independent | Source ↗ |
| MMLU-Redux | Saturated | 86.2% | 25 | 2026-08-25 | independent | Source ↗ |
| QuALITY | Nearing Saturation | 77.4% | 78 | 2026-08-25 | independent | Source ↗ |
| LongBench | Nearing Saturation | 75.7% | 65 | 2026-08-24 | independent | Source ↗ |
| SWE-bench | Deprecated | 51.3% | 10 | 2026-08-24 | independent | Source ↗ |
| MGSM | Nearing Saturation | 81.7% | 65 | 2026-08-24 | independent | Source ↗ |
| AGIEval | Saturated | 74.4% | 25 | 2026-08-23 | independent | Source ↗ |
| QASPER | Nearing Saturation | 73% | 65 | 2026-08-23 | independent | Source ↗ |
| TruthfulQA | Saturated | 77.4% | 25 | 2026-08-23 | independent | Source ↗ |
| GSM-Hard | Deprecated | 82.6% | 10 | 2026-08-22 | independent | Source ↗ |
| SuperGLUE | Saturated | 88.2% | 25 | 2026-08-22 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 71% | 23 | 2026-08-21 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 6.4% | 75 | 2026-08-21 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 84.8% | 25 | 2026-08-21 | independent | Source ↗ |
| WinoGrande | Saturated | 90% | 28 | 2026-08-20 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 77% | 65 | 2026-08-20 | independent | Source ↗ |
| IFEval | Deprecated | 60.8% | 25 | 2026-08-20 | independent | Source ↗ |
| GPQA | Nearing Saturation | 77.4% | 78 | 2026-08-20 | independent | Source ↗ |
| Needle-in-a-Haystack | Saturated | 97.6% | 44 | 2026-08-19 | independent | Source ↗ |
| HumanEval+ | Saturated | 81.5% | 25 | 2026-08-19 | independent | Source ↗ |
| CommonsenseQA | Saturated | 90.8% | 28 | 2026-08-19 | independent | Source ↗ |
| MBPP | Saturated | 80% | 25 | 2026-08-18 | independent | Source ↗ |
| GSM8K | Saturated | 92.9% | 25 | 2026-08-18 | independent | Source ↗ |
| GLUE | Saturated | 93.4% | 28 | 2026-08-18 | independent | Source ↗ |
| MMLU | Saturated | 89.8% | 25 | 2026-08-17 | independent | Source ↗ |
| HumanEval | Saturated | 86.1% | 25 | 2026-08-17 | independent | Source ↗ |
| ARC-AGI | Saturated | 81.3% | 25 | 2026-08-17 | independent | Source ↗ |
| LiveBench (Continuous) | Active | 74.5% | 94 | 2025-02-01 | independent | Source ↗ |
Local Hardware Execution (Ollama)Hardware Compatible
# 1. Pull and execute model locally with Ollama
ollama run deepseek-v4-flash
# 2. Or invoke via local OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-flash",
"messages": [{"role": "user", "content": "Analyze reasoning chains on AIME 2026"}]
}'Evaluation & Sourcing Notes
Scores listed for DeepSeek V4-Flash represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.
All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.