DeepSeek V4-Flash 0731 Tops V4-Pro at a Third the Price

DeepSeek shipped something on July 31 that is easy to underestimate: the same model, made substantially better without touching the architecture. DeepSeek-V4-Flash-0731 is the April preview rebuilt with a much stronger post-training pipeline, and the result is a model that outperforms the company's own larger V4-Pro preview on every agentic benchmark DeepSeek published — while costing roughly a third as much per output token.

• Same architecture, new post-training: 284B total parameters, 13B active per token, 256 routed experts per layer, 1M-token context
• DeepSWE jumped from 7.3 to 54.4 and Cybergym from 38.7 to 76.7 versus the April preview
• Pricing: $0.14 per million input tokens on a cache miss, $0.28 per million output tokens
• MIT licensed and ungated, self-hostable in roughly 110 GB at 3-bit quantization

What Actually Changed in the 0731 Build?

Nothing in the model's skeleton. The parameter count, the mixture-of-experts routing, the hybrid compressed sparse attention and the million-token context window are all carried over from the April preview we covered when DeepSeek's V4 family first landed. Every gain in the 0731 build comes from retraining the same weights through a redesigned post-training stage aimed squarely at agentic workflows: long tool-use chains, multi-step reasoning, and code that has to survive a real terminal.

That is a meaningful data point for anyone tracking where the open-weight frontier is heading. For most of 2025 and early 2026, the headline lever was scale. The 0731 build suggests a second lever is now doing serious work — the quality of what happens after pretraining. It is also the cheaper lever, which is exactly why the output price fell rather than rose.

How Big Are the Agentic Gains?

The two most striking published deltas are software-engineering and security benchmarks. On DeepSWE, the agentic software-engineering suite, the score moved from 7.3 to 54.4. On Cybergym, a benchmark for autonomous vulnerability reasoning, it went from 38.7 to 76.7. DeepSeek also reports a Terminal Bench 2.1 score of 82.7, with native support for the Responses API and Codex-style tooling.

A fair caveat, and DeepSeek's own documentation is upfront about it: these are vendor-reported numbers measured on a harness that has not been released. Benchmark jumps of that magnitude are the kind of claim that deserves your own evaluation run before you rewire a production pipeline around it. The encouraging part is that the model is MIT licensed and ungated, so running that evaluation costs you compute rather than a procurement cycle.

What Does V4-Flash 0731 Cost to Run Locally?

This is where the mixture-of-experts design earns its keep. Only 13B of the 284B parameters activate per token, so inference is far lighter than the total parameter count implies. DeepSeek puts self-hosting at roughly 110 GB of memory at 3-bit quantization, or a 4×GB300 node if you want full precision and the model's 384K-token output ceiling at higher reasoning effort.

110 GB is genuinely reachable — it is a well-specified workstation or a modest multi-GPU box rather than a data center. If you are weighing quantization levels for a build like that, our LLM quantization guide comparing GGUF, AWQ and MLX walks through the trade-offs that decide whether 3-bit is usable for your workload or a step too far.

Why This Matters for Open-Weight AI

2026 has been a remarkable year for open weights. Kimi K3 shipped 2.8 trillion parameters with a million-token context, and now DeepSeek has demonstrated that a mid-sized open model can leapfrog its own larger sibling through better training craft alone. For teams building agents, that combination — permissive licensing, million-token context, sub-$0.30 output pricing, and a self-hostable footprint — is about as friendly a starting point as the AI landscape has offered.

The practical advice is unchanged and worth repeating: benchmark it yourself on your own tasks. But the direction of travel is hard to miss.

Sources: MarkTechPost — July 31, 2026; Simon Willison — July 31, 2026; TechTimes — July 31, 2026.