DeepSeek V4 Flash 0731: How Re-Post-Training Beats Scaling
DeepSeek V4 Flash 0731 demonstrates how focused re-post-training enables a 13B active parameter MoE model to surpass larger architectures across key agentic and coding benchmarks.
DeepSeek V4 Flash 0731: How Re-Post-Training Beats Scaling
A 13B active parameter MoE model outperforming heavier architectures on agentic tasks.
On July 31, 2026, DeepSeek released V4-Flash-0731 via their API changelog and open-sourced the weights on Hugging Face under the MIT license. Rather than scaling pre-training parameters, DeepSeek kept the architecture identical to the previous V4-Flash preview and focused entirely on re-post-training.
The 13-billion-active-parameter model surpasses its 49-billion-active-parameter Pro sibling in agentic tasks, showing how post-training can outperform sheer parameter scale.
Architecture: Sparse Activation
As a 284-billion parameter Mixture of Experts (MoE) model, DeepSeek V4 Flash 0731 activates only 13 billion parameters per token. This sparse activation keeps inference fast and cost-effective.
| Specification | Value |
|---|---|
| Total Parameters | 284 Billion |
| Active Parameters (per token) | 13 Billion |
| Architecture | MoE (Mixture of Experts) |
| Context Window | 1 Million Tokens |
| License | MIT (Open Weights) |
The model includes the DSpark speculative-decoding module for high inference throughput, though developers must apply specific weight patches to prevent throughput from dropping by up to 50%.
Benchmark Performance & Point Improvements
Official evaluation data published on Hugging Face and DeepInfra details gains across all nine core benchmarks compared to the V4-Flash Preview and V4-Pro Preview:
| Benchmark | V4 Flash 0731 | V4 Flash Preview | Point Gain | V4 Pro Preview | GLM-5.2 | Opus-4.8 |
|---|---|---|---|---|---|---|
| DeepSWE | 54.4 | 7.3 | +47.1 | 12.8 | 46.2 | 58.0 |
| Terminal Bench 2.1 | 82.7 | 61.8 | +20.9 | 72.1 | 81.0 | 85.0 |
| Cybergym | 76.7 | 38.7 | +38.0 | 52.7 | — | 83.1 |
| Toolathlon Verified | 70.3 | 49.7 | +20.6 | 55.9 | 59.9 | 76.2 |
| DSBench-FullStack (Internal) | 68.7 | 37.0 | +31.7 | 41.8 | 61.8 | 71.6 |
| DSBench-Hard (Internal Agent) | 59.6 | 25.8 | +33.8 | 31.1 | 54.5 | 71.7 |
| NL2Repo | 54.2 | 39.4 | +14.8 | 38.5 | 48.9 | 69.7 |
| Agents' Last Exam | 25.2 | 15.8 | +9.4 | 16.5 | 23.8 | 25.7 |
| AutomationBench Public | 25.1 | 10.8 | +14.3 | 12.8 | 12.9 | 27.2 |
Specific Gains
- DeepSWE: Rose to 54.4 from 7.3 (+47.1 points), outperforming open models like GLM-5.2 (46.2) and coming within 3.6 points of proprietary frontier models like Opus-4.8 (58.0).
- Terminal Bench 2.1: Reached 82.7 (+20.9 points over preview), placing within 0.6 points of proprietary models like Grok 4.5 (83.3).
- Cybersecurity & Tool Use: Increased by 38.0 points on Cybergym (to 76.7) and 20.6 points on Toolathlon Verified (to 70.3).
Note: V4-Pro preview scored 67.9 on Terminal Bench 2.0, while 0731 scored 82.7 on Terminal Bench 2.1. Because these use different benchmark versions, direct comparisons require version alignment.
Comparison with Leading Open-Source Models
Independent testing by Artificial Analysis highlights how V4 Flash 0731 ranks among open-weights models:
1. Real-World Agent Tasks (GDPval-AA v2 Elo)
On GDPval-AA v2, an evaluation focused on real-world work tasks, V4 Flash 0731 gained 370 Elo points over its predecessor:
| Model | License / Weight Access | Elo Rating |
|---|---|---|
| Kimi K3 (max) | Open Weights | 1687 |
| DeepSeek V4 Flash 0731 | Open Weights | 1559 |
| GLM-5.2 (max) | Open Weights | 1510 |
| DeepSeek V4 Flash (Previous) | Open Weights | 1189 |
V4 Flash 0731 holds the second-highest open-weights Elo rating on GDPval-AA v2, trailing Kimi K3 (max) while leading GLM-5.2 by 49 Elo points.
2. Intelligence Index Gains
On the Artificial Analysis Intelligence Index, V4 Flash 0731 scored 50 (+10 points over the April release, +6 points over V4-Pro):
- CritPt: +9 points (to 17%)
- SciCode: +5 points (to 50%)
- Humanity's Last Exam: +5 points (to 37%)
- AA-LCR: +3 points (to 66%)
- GPQA Diamond: +1 point (to 91%)
3. Output Efficiency and Hallucinations
- Token Efficiency: Used 12% fewer output tokens (2.06B vs. 2.34B) during benchmark runs.
- Hallucinations: AA-Omniscience Index score improved by 7 points, driven by a 12-percentage-point decrease in hallucinations (down to 84%), comparable to models like GPT-5.6 Terra max (85%) and Mistral Medium 3.5 (82%).
Pricing and Efficiency
DeepSeek maintains the same pricing structure:
- Input: $0.14 per 1M tokens
- Output: $0.28 per 1M tokens
- Cache Hit: $0.003 per 1M tokens (~98% discount)
Running a complex agentic task on DeepSeek V4 Flash 0731 is approximately 60% cheaper than running it on GPT-5.6 Luna (max), putting it on the Pareto frontier for intelligence versus cost per task in Artificial Analysis rankings.
Additional Updates
- Native Responses API Support: The update adds native support for the Responses API format, enabling integration with Codex CLI, the ChatGPT desktop app, and the VS Code Codex plugin.
- DeepSeek Harness: Benchmark scores were generated using the unreleased "minimal mode" of DeepSeek's Harness framework. The reported results reflect system-level performance (model, reasoning budget, and framework) rather than weight-only inference.
Limitations and Deployment Notes
- Evaluation Environment: Because DeepSeek Harness is unreleased and some benchmarks use internal datasets, third parties cannot yet fully replicate all official scores.
- Throughput Patches: Deploying the open weights without applying required throughput patches can reduce decoding speed by up to 50%.
- Consumer Interfaces: The updated build is live on the API and Hugging Face, but has not yet rolled out to the consumer web or mobile applications.
Conclusion
DeepSeek V4 Flash 0731 demonstrates that refined post-training can yield larger capability gains than increasing active parameter counts. By keeping a 13B active parameter architecture and retraining post-pre-training pipeline stages, DeepSeek produced a model that outperforms its 49B active parameter sibling on key agentic benchmarks.
DeepSeek plans to release a similar post-training update for V4-Pro in early August.
V4-Flash-0731 is available via the DeepSeek API (model="deepseek-v4-flash") and on Hugging Face at deepseek-ai/DeepSeek-V4-Flash-0731.