DeepSeek releases V4.1-Flash, the smallest model in its new architecture with native multimodal vision

On 10 September 2026 DeepSeek published an update to its API changelog announcing DeepSeek-V4.1-Flash, presented as the smallest model within a new architecture family. According to the company's own documentation, this design is intended to deliver "a higher capability ceiling, faster inference, greater throughput and scalability towards larger models", as well as incorporating native multimodal visual understanding within the model itself, without relying on a separate experimental component as was the case with the previous version.

The document accompanies the launch with an extensive battery of benchmark results, all quoted verbatim from the source: GPQA Diamond (90.9), HLE (36.8, or 39.1 if measured only on the pure-text subset), Codeforces with a rating of 3471, MathArena Apex (65.6), and a series of tests focused on agents and terminal tasks: Terminal-Bench 2.1 (90.6), Terminal-Bench 3.0 (30.0) and Terminal-Bench 4.0 (31.2). Figures are also reported for more specialised benchmarks such as DeepSWE v1.1 (74.2), ProgramBench (20.3), NL2Repo-Bench (65.4), CyberGym (88.1), SEC-Bench Pro (62.8) and ExploitGym (15.3), the latter aimed at evaluating offensive/defensive security capabilities and vulnerability exploitation in controlled environments. On the front of tool-using agents, DeepSeek reports HLE with tools (63.9), Automation-Bench (54.8), Agents' Last Exam (31.8), Chartography with tools (78.9), BabyVision with tools (89.6) and ZeroBench-main with tools (49.0). This is a broad and heterogeneous set of metrics, spanning everything from academic reasoning to automation and cybersecurity tasks, although the document offers no direct comparisons against competing models for this particular version, so these figures are best taken as DeepSeek's own internal reference data rather than as an externally validated comparison.

On the operational side, the most relevant change for developers is that the model name deepseek-flash now points to this new V4.1 Flash version. The previous versions, V4 Flash and V4 Flash Vision Exp, are being retired, although for compatibility the old names (deepseek-v4-flash and deepseek-v4-flash-vision-exp) will temporarily continue routing to V4.1 Flash, giving teams integrating the API room to migrate without abrupt disruption. In addition, in response to user demand, DeepSeek has decided to keep the API service for DeepSeek V4 Pro running beyond 14 September 2026, retaining the same billing method, though it warns there could be changes in the future. Alongside the launch, the company has announced a price reduction for the API, pointing to its pricing page for the exact details.

The announcement comes amid a very intense update cadence from DeepSeek throughout 2026, which the changelog itself documents step by step. Just three weeks earlier, on 21 August, the company had launched DeepSeek-V4-Flash-Vision-Exp, an experimental visual understanding model accessible via the parameter model='deepseek-v4-flash-vision-exp'. In its own testing, that experimental model matched the official V4-Flash version on pure-text capabilities (agent, reasoning, general knowledge), but showed a notable jump on agent benchmarks requiring visual understanding, bringing —according to DeepSeek itself— its multimodal agent capabilities close to those of Opus-4.8. That is a claim made by the company in its own documentation, with no independent verification cited in the text.

A few weeks before that, on 13 August, DeepSeek-V4-Pro had reached general availability (GA) on app, web and API, with improvements in agent capabilities described as "particularly significant in production environments", and with native support for OpenAI's Responses API format, adapted specifically for Codex. That same update introduced three reasoning effort levels (low, high and max) for both V4-Pro and V4-Flash, designed to let users adjust compute consumption according to task complexity: low for simple tasks, high for everyday agent tasks and max for more complex scenarios. It also introduced a peak/off-peak pricing scheme at that point, with off-peak rates set at half the peak-hour price, as a way of encouraging a more efficient distribution of usage load.

Going back a little further, the official beta of DeepSeek-V4-Flash had been made public on 31 July, with agent improvements the company describes as far superior to the preliminary V4-Pro-Preview version, and with the same architecture and size as its predecessor, albeit with different post-training. The V4 family as a whole had debuted on 24 April 2026, gradually replacing the legacy names deepseek-chat and deepseek-reasoner, whose discontinuation was scheduled for three months later.

Taken together, the changelog sketches a strategy of very rapid iteration on DeepSeek's part: in just under five months the company has gone from introducing the V4 architecture to launching a Pro GA version, a Flash beta and official release, an experimental vision model and, finally, a 4.1 Flash version that unifies and replaces the previous ones with native multimodal support out of the box. For teams integrating the API, the practical message is clear: the old model names will keep working for now via redirection, but it is advisable to migrate to the current names (deepseek-flash, deepseek-v4-pro) and review the pricing documentation, given that the rate policy has changed several times over these months, including the introduction of time-of-day differentiated pricing and the latest reduction tied to the V4.1 Flash launch.