DeepSeek has achieved a major jump in artificial intelligence agent performance without altering its model's underlying structure, reaching 82.7 on Terminal-Bench 2.1 through post-training improvements alone. The company released DeepSeek-V4-Flash-0731 as a public beta through its API on July 31, 2026, publishing open weights on Hugging Face under the MIT license the same day. The update demonstrates that meaningful gains in AI capability can come from refining existing models rather than building bigger ones.
The updated Flash version maintains the same architecture as the preview release, operating with 284 billion total parameters and 13 billion activated parameters per token. This stands far smaller than V4-Pro, which runs 1.6 trillion total parameters and 49 billion activated parameters. Despite being the more compact model, the new Flash iteration outperformed the earlier V4-Pro preview across multiple agent-focused benchmarks, posting 54.4 on DeepSWE and 70.3 on Toolathlon-Verified alongside its Terminal-Bench result. Early independent testing by Artificial Analysis found a slightly lower Terminal-Bench 2.1 score of 79%, suggesting DeepSeek's reported figures may not always align with outside verification.
According to the company, additional post-training drives the performance improvements, proving that substantial advances don't always demand a larger model. DeepSeek also shared outcomes from several internal tests, though independent verification hasn't yet confirmed them. The release includes support for the Responses API for constructing AI agents and multi-step workflows, with published instructions for weaving the model into Codex-based development processes.
The MIT license gives organizations full control over deployment and customization, freeing them from reliance on DeepSeek's hosted API. For companies running agents at scale, the gap in activated parameters translates directly into inference costs, though model pricing alone doesn't capture the complete picture. The V4 technical report details inference optimizations like speculative decoding with DeepSeek's DSpark framework, designed to make serving more efficient. When paired with self-hosted deployments, these features let infrastructure teams fine-tune performance for their own production requirements. Support for OpenAI-style APIs lowers the barrier for teams already using those formats, allowing them to evaluate the model without overhauling existing setups.
The release signals a shift in how companies extract better performance from AI models, focusing on post-training refinement instead of simply expanding size. While many AI vendors concentrate on hosted APIs, DeepSeek continues publishing downloadable weights that organizations can run on their own infrastructure. The move adds to a broader trend of open-weight models closing the performance gap with proprietary alternatives, competing increasingly on delivery methods rather than just raw capability. Organizations now have more ways to boost model effectiveness without the infrastructure costs and complexity that come with larger architectures.