DeepSeek-V4-Flash-0731 beats V4-Pro on all nine agent benchmarks

DeepSeek-V4-Flash-0731, the official release of DeepSeek's budget-tier Mixture-of-Experts model, outperforms the company's own V4-Pro (Preview) on all nine agent benchmarks in its evaluation set. The build shipped July 31, 2026 as a re-post-training of the preview, keeping the architecture and size unchanged while reworking the training pipeline around coding, AI agents, reasoning, and tool use. API pricing stays at $0.14 per million input tokens.

The scoreboard is where this release earns attention. On Terminal Bench 2.1, which scores agent behavior in terminal-driven coding environments, the model lands at 82.7 against 61.8 for the preview. On DeepSWE, an end-to-end software engineering benchmark, it climbs from 7.3 to 54.4, a jump of roughly 645%. Both results also beat the larger V4-Pro (Preview), an unusual outcome for a budget line: the cheap model is now the strongest option for agentic coding, not a compromise.

Because the architecture is identical to the preview, the gains are purely a post-training story. DeepSeek re-post-trained the weights with an improved pipeline centered on coding, agents, reasoning, and tool use, and the release describes the agentic capability as substantially stronger than the preview's. What stands out to me is that the entire nine-benchmark sweep moved in the same direction, which suggests the training data and pipeline are doing the work rather than raw parameter count.

What DeepSeek-V4-Flash-0731 Changes for Agentic Workflows

The architecture carries over from the preview: a roughly 300-billion-parameter MoE that activates 13 billion parameters per token, with the DSpark speculative decoding module keeping generation fast enough for interactive agent loops. The new reasoning_effort parameter is the headline control, offering three levels (low, high, max) so teams can dial reasoning depth per task. The trade-off is direct: higher effort consumes more output tokens and latency, which is why the model card recommends a 384K max output window only for the high and max settings.

Context support reaches one million tokens, enough for extended conversation histories, tool outputs, and repository context inside a single session. That matters for agentic coding, where one run can accumulate a long working transcript before it emits a patch. Self-hosting is where the details get interesting: DeepSeek-V4-Flash-0731 ships without a Jinja chat template and relies on a dedicated Python-based encoding system for OpenAI-compatible formatting, so local integrations should plan around that path rather than a stock template.

Weights are published in BF16, FP8, and MXFP4 precisions under the MIT license, which permits commercial use and modification. DeepSeek points to LM Studio and vLLM recipes for local deployment, and the 0731 build is already the default DeepSeek-V4-Flash in the vLLM recipe collection, so the standard serving stack targets it out of the box. Because the serving format is OpenAI-compatible, existing agent frameworks that already speak that protocol can point at the model with minimal changes to their prompt plumbing.

For teams already running the preview, the switch is nearly free: same architecture, same price, and the same OpenAI-compatible serving format, so moving to the official release is a weight swap rather than a migration. API users do not even change endpoints, and local users just pull the new 0731 weights in a preferred precision.

The pricing compounds the impact. The API rate is unchanged at $0.14 per million input tokens, so the agentic gains arrive without a cost increase. At that rate, even a context-heavy agent run that consumes half a million input tokens costs well under a dime on inputs, which makes long retry loops and exploratory agent passes affordable at scale. The effective capability per dollar on DeepSWE rose by roughly the same factor as the score itself.

Why This Matters

DeepSeek-V4-Flash-0731 collapses the usual tier logic: the budget model is now the first choice for agentic coding pipelines, and V4-Pro has to justify its premium on other grounds. The main setup wrinkle for local use is the custom Python encoding path, since there is no stock chat template, so budget a little integration time there. The practical move for most agent builds is to point the pipeline at the 0731 build with high reasoning effort and measure the results on your own workloads rather than trusting benchmark sheets alone.

Related Articles

• DeepSeek V4 Launch Introduces Trillion-Parameter Pro and High-Speed Flash Models
• DeepSeek V4 Pro Adoption Surges as US Firms Pivot to Low-Cost AI Models
• DeepSeek DSpark Framework Speeds LLM Inference 85% in Open-Source Push

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.