DeepSeek formally released V4.1 Flash on September 10, a 552-billion-parameter Mixture-of-Experts model on a new Causal-Encoder-Decoder architecture that the company says surpasses its previous flagship, V4 Pro, in performance, cost, speed and total completion time2,5,7. Starting September 14 at noon Beijing time, all requests to the V4 Pro endpoint will be routed to V4.1 Flash and billed at Flash-series rates, effectively retiring the larger model.
The release followed a two-day limited beta that began around September 8 under a time-stamped model identifier, deepseek-v4.1-flash-expires-on-0910, capped at 20 concurrent requests per account1,6.
Asymmetric activation and cache compression
V4.1 Flash uses asymmetric activation: roughly 8 billion parameters fire on the input side and roughly 16 billion on the output side11. DeepSeek argues the split cuts compute cost relative to denser models of similar total size, a design it says suits agent workloads that ingest large codebases and documents before emitting shorter decisions or tool calls. The model supports contexts of up to one million tokens10,12.
KV-cache compression is a central claim. DeepSeek says HBM requirements drop to one-quarter and SSD requirements to one-eighth compared with the previous generation, and that the KV cache has shrunk 437 times relative to the first-generation model. The headline from MarkTechPost references FP4 KV cache and cross-layer attention reuse as contributing techniques8.
Pricing and multimodal consolidation
New Flash-series pricing took effect at noon Beijing time on September 10. Off-peak rates are RMB 0.02 per million tokens for cache-hit input, RMB 1 for cache-miss input and RMB 4 for output; peak-hour prices are double. Off-peak cache-hit input was cut about 60 percent versus prior Flash pricing.
Native vision understanding now lives inside the primary Flash API rather than through the earlier bolt-on vision experiment. Older V4 Flash and V4 Flash Vision Exp endpoints are offline; for backward compatibility, those model identifiers temporarily route to V4.1 Flash. On the consumer app, DeepSeek's Quick, Expert and Vision modes merge into a single smart mode that detects complexity, enables vision when images arrive, and scales thinking without manual switching.
Benchmark claims and competitive framing
DeepSeek said internal and external testing showed V4.1 Flash ahead of V4 Pro across evaluated suites, citing results on GPQA Diamond, Codeforces-style programming, MathArena Apex, Terminal-Bench 2.1 and CyberGym. An IndexBox headline placed V4.1 Flash ahead of GPT-5.6 Sol and Kimi K33, and the South China Morning Post reported DeepSeek's claim that the model beats Kimi K3 on cyber and coding benchmarks4,9. The model is described as open-weight and has been posted to Hugging Face13.
Tencent's WorkBuddy and CodeBuddy products and OpenCode have integrated V4.1 Flash as launch partners.
ANALYSIS The move collapses DeepSeek's product line: a smaller, cheaper model absorbs the flagship's traffic on a fixed calendar, rather than coexisting alongside it. The asymmetric activation design and aggressive cache compression target the cost profile of long-context agent pipelines, where input tokens vastly outnumber output tokens.