DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains

DeepSeek has officially launched DeepSeek-V4-Flash-0731, marking a major upgrade from its preview. The model, now in public beta via API, shows strong improvements in agent and coding benchmarks through additional training. This release gives developers a more affordable and capable option for demanding AI work.

Key takeaways

• DeepSeek-V4-Flash-0731 is now publicly available via API and for self-hosting, superseding the preview version.
• The model achieves major performance gains in agent and coding tasks through re-post-training, not architectural changes.
• API pricing offers output tokens at roughly one-third the cost of the V4-Pro model, improving affordability.
• Self-hosting requires significant resources, with options for full-precision on a 4×GB300 node or a 3-bit quantized build needing ~110 GB RAM/VRAM.
• All reported benchmark scores are vendor-generated using an unreleased harness, highlighting the need for independent evaluation.

Release Details and Deployment Options

DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API into public beta on July 31, 2026. The model card confirms this release supersedes the preview, with gains from re-post-training rather than a new architecture.

You can deploy it via API or self-hosting. API pricing is listed at $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens. This output cost is about a third of the deepseek-v4-pro output price of $0.87.

For self-hosting, the model is MIT-licensed and ungated, but needs substantial resources. DeepSeek’s vLLM example serves it on a single 4×GB300 node for full-precision. A 3-bit quantized build requires around 110 GB of combined RAM plus VRAM.

Architectural Specifications and Efficiency

DeepSeek-V4-Flash-0731 is a 284-billion parameter Mixture of Experts (MoE) model. It activates about 13 billion parameters per token and supports a 1-million token context window. Its technical report states each MoE layer contains 1 shared expert and 256 routed experts, with 6 routed experts firing per token. The model uses a hybrid attention mechanism that blends Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).

Instead of standard residual connections, the architecture employs Manifold-Constrained Hyper-Connections (mHC). It has an expansion factor of 4 and uses 20 Sinkhorn-Knopp iterations. For serving, the model checkpoint ships with the DSpark speculative decoding module attached. You can enable this module with a specific vLLM flag. The associated paper claims it provides 60–85% faster per-user generation than a baseline model.

Reported Performance Benchmarks

DeepSeek’s published model card states the new DeepSeek-V4-Flash-0731 beats both the V4-Flash (Preview) and V4-Pro (Preview) models on every listed agent and coding benchmark. The reported scores show clear gains.

Key Benchmark Scores

The vendor-reported benchmarks for V4-Flash-0731 include:

• Terminal Bench 2.1: 82.7
• NL2Repo: 54.2
• Cybergym: 76.7
• DeepSWE: 54.4
• Toolathlon-Verified: 70.3
• Agents’ Last Exam: 25.2
• AutomationBench Public: —

These scores are all DeepSeek-reported figures. The code agent tasks used the minimal mode of the DeepSeek Harness, which hasn’t been publicly released. The source notes that agent scores are harness-sensitive. It suggests running independent evaluations, as results could differ.

Key Takeaways and Practical Considerations

The performance gains in DeepSeek-V4-Flash-0731 come from re-post-training, not a change to the core 284B/13B MoE architecture from the April preview. The model natively supports the Responses API format and is adapted for Codex.

Serving and Deployment

For agent use, DeepSeek suggests configuring the API with temperature=1.0 and top_p=0.95. Self-hosting the MIT-licensed model demands significant resources. A 3-bit quantized build needs around 110 GB of combined RAM and VRAM, or use a 4×GB300 node for full-precision serving.

Evaluation Context

All published benchmark numbers, which show the new model beating the V4-Pro preview on agent tasks, are vendor-reported using an unreleased harness (the minimal mode of DeepSeek Harness). Since agent scores are harness-sensitive, you should seek independent evaluation before production use.