DeepSeek V4.1 Flash (Consensus Protocol)

A third-party review published 2026-09-10 characterizes DeepSeek-V4.1-Flash as the newest Flash model from DeepSeek, designed for higher capability ceiling, faster inference, higher throughput, and scaling to larger models, with native multimodal visual understanding integrated into the new architecture rather than add The same review notes that published coding results improve sharply versus the previous V4 Flash 0731 release, with Terminal-Bench 2.1 rising from 82.7 to 90.6 and DeepSWE from 54.4 to 74.2. It reports DeepSeek's Flash pricing effective 2026-09-10 at $0.003 per million cache-hit input tokens, $0.15 per million uncached

A third-party technical analysis published on 2026-09-10 describes DeepSeek V4.1 Flash as introducing a Causal Encoder-Decoder (CED) architecture with a 552B-parameter backbone, activating roughly 8B parameters during prefill and 16B during decode, and claims a roughly 4× KV cache reduction versus V4 Flash and 437× ver The same analysis reports that V4.1 Flash outperforms DeepSeek V4 Pro on all measured benchmarks despite roughly one-third the total parameters and one-sixth the active parameters, framing this as the first time a Flash-tier model fully replaces a Pro-tier model in the DeepSeek lineup. It positions the gains as an effi

A Substack analysis dated 2026-09-15 traces DeepSeek's KV cache compression trajectory, noting that DeepSeek-V4 (April 2026) introduced CSA and HCA techniques that reduced KV cache storage by roughly 4× and 128× respectively, and built them atop sparse attention with a top-k of 512 indexer to further cut compute during The same article positions V4.1 as a further step beyond V4, achieving approximately 3.9× additional KV cache compression on top of an already compressed state, extending the KV-reduction storyline that DeepSeek has pursued over the last three years. The analysis discusses these compression techniques in the context of

DeepSeek's official API changelog for 2026-09-10 documents the release of DeepSeek-V4.1-Flash as the smallest model in the company's new architecture family, with native multimodal visual understanding, a design intended for higher capability, faster inference, higher throughput, and scaling to larger models. The chang The same changelog entry confirms that DeepSeek V4.1 Flash is available on the DeepSeek API with native multimodal support, and that the new canonical model name is 'deepseek-flash', which routes to V4.1 Flash, while the previous V4 Flash and V4 Flash Vision Exp have been retired (their legacy names are temporarily ali