DeepSeek-V4-Flash-Max
DeepSeek-V4-Flash-Max is the maximum reasoning effort mode of DeepSeek-V4-Flash, a 284B-parameter MoE model with 13B activated parameters and a 1M-token context window. Sharing the V4 series' hybrid attention architecture (Compressed Sparse Attention combined with Heavily Compressed Attention), Manifold-Constrained Hyper-Connections, and Muon optimizer, V4-Flash-Max delivers reasoning performance comparable to V4-Pro when given a larger thinking budget while operating at a fraction of the parameter scale. It is pre-trained on more than 32T tokens and post-trained with a two-stage paradigm of domain-specific expert cultivation followed by on-policy distillation.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| BrowseComp | 73.2% | self-reported llm-stats | link → |
| CodeForces | 100.0% | self-reported llm-stats | link → |
| CorpusQA 1M | 60.5% | self-reported llm-stats | link → |
| CSimpleQA | 78.9% | self-reported llm-stats | link → |
| GDPval-AA | 1,395 | self-reported llm-stats | link → |
| GPQA | 88.1% | self-reported llm-stats | link → |
| HMMT Feb 26 | 94.8% | self-reported llm-stats | link → |
| Humanity's Last Exam | 45.1% | self-reported llm-stats | link → |
| IMO-AnswerBench | 88.4% | self-reported llm-stats | link → |
| LiveCodeBench | 91.6% | self-reported llm-stats | link → |
| MathArena Apex | 85.7% | self-reported llm-stats | link → |
| MCP Atlas | 69.0% | self-reported llm-stats | link → |
| MMLU-Pro | 86.2% | self-reported llm-stats | link → |
| MRCR 1M | 78.7% | self-reported llm-stats | link → |
| SimpleQA | 34.1% | self-reported llm-stats | link → |
| SWE-bench Multilingual | 73.3% | self-reported llm-stats | link → |
| SWE-Bench Pro | 52.6% | self-reported llm-stats | link → |
| SWE-Bench Verified | 79.0% | self-reported llm-stats | link → |
| Terminal-Bench 2.0 | 56.9% | self-reported llm-stats | link → |
| Toolathlon | 47.8% | self-reported llm-stats | link → |