DeepSeek releases V4.1-Flash with causal encoder-decoder, 1M-token context and vision

The headline number is counterintuitive: at 763B parameters V4.1-Flash is larger than the V3 and R1 models that made DeepSeek famous, yet its memory footprint is far lower than its size implies. DeepSeek achieved this through overhauled attention mechanisms and a new causal encoder-decoder (CED) that improves prompt processing while slashing KV-cache use to 13-25% of V4 Flash's — meaning the same hardware can serve four to eight times more concurrent users.

The most novel change is architectural: 196B of the 763B parameters are 'N-gram parameters' forming what DeepSeek calls a 'conditional memory module.' By decoupling memory from computation, DeepSeek argues it can make models smarter while reducing the compute and memory needed to serve them — a template The Register suggests could point toward larger, leaner future models. The model adds vision and a 1M-token context, and DeepSeek claims it outperforms flagship V4-Pro and beats Kimi K3 on cyber and coding benchmarks.

Pricing is the competitive weapon. Off-peak rates of $0.15/$0.60 per million tokens, with low cached-input costs, drew a claimed ~94% cost reduction versus GPT-5.6 Sol from @kimmonismus, who also cited it beating Sol on DeepSWE (74.2% vs 73%). OpenCode integrated day one; open weights run on Ollama, LM Studio and llama.cpp.

The launch is inseparable from Anthropic's same-week accusation that DeepSeek distilled Claude — a charge the community debated directly on r/DeepSeek. It also arrives ahead of a Shanghai STAR Market listing, adding a financial motive to the aggressive pricing. V4 Pro will be retired September 14, though DeepSeek later signaled it would continue supporting it. Watch third-party benchmark reproductions to confirm the V4-Pro-beating claims.