DeepSeek has shipped V4.1-Flash-1017, a dated build of its V4.1 Flash model, scoring 77.0 on DeepSWE, the firm's software-engineering benchmark, and opening up a search API alongside it. The naming convention, with the build date baked into the model identifier, is DeepSeek's usual way of marking an incremental update rather than a full version bump.
V4.1 Flash itself is a mixture-of-experts model, with reporting around its launch putting total parameters at roughly 552 billion and active parameters in the single-digit billions per token, which keeps inference cost down relative to a dense model of similar scale. It supports a context window stretching past the million-token mark, and earlier V4.1 Flash releases already outperformed the prior V4-Flash generation across benchmarks including Terminal-Bench and CyberGym.
The new piece here is the search API. Opening a dedicated search capability lets developers call out to live web results directly through DeepSeek's own infrastructure during inference, rather than bolting on a third-party retrieval layer. Combined with the DeepSWE score, the update reads as DeepSeek sharpening V4.1 Flash specifically for agentic coding work, where a model needs to both reason about code and pull in up-to-date information mid-task.
DeepSeek has kept up a fast release cadence through 2026, iterating on the Flash line in frequent, dated increments rather than infrequent major versions. For developers already building on the V4.1 Flash API, the practical change is a model ID update and access to the new search endpoint, with no indication of a pricing change alongside this release.
Source: MarkTechPost
