DeepSeek V4.1-Flash launch

V4.1-Flash's headline number is deceptive. The model carries a 552B-parameter backbone plus 196B additional Engram parameters for a 763B total^[1], but a new Causal-Encoder-Decoder design - a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder - projects the decoder's global KV cache from the encoder's final hidden states instead of recomputing it at every decoder layer^[1]. That lets the model activate just 8B parameters per token during prefill and 16B during decode^[1], regardless of how large the underlying weights are. Paired with FP4 KV caching (E2M1 format) and Compressed Sparse Attention 2, the global KV cache footprint drops to 890 bytes per token - about a quarter of DeepSeek-V4-Flash's footprint and roughly 437x smaller than DeepSeek-V1's^[2]. The context window stretches to 1 million tokens, and DeepSeek says scaling from 4K to 1M tokens adds only about 25% extra computation^[3]- a sign the architecture, not just brute-force compute, is doing the work.

DeepSeek is making explicit benchmark claims against the two reigning closed-model leaders. On Terminal-Bench 2.1, V4.1-Flash scores 90.6 versus 88.8 for GPT-5.6 Sol and 89.1 for Claude Opus-5; on DeepSWE v1.1 it scores 74.2 versus 73.0 and 74.0 respectively^[2]. Those are narrow margins, but they arrive alongside off-peak API pricing of RMB 0.02 per million cache-hit input tokens, RMB 1 for cache-miss input, and RMB 4 for output (peak-hour rates double)^[4]- a fraction of what frontier-lab pricing typically runs. The model isn't only chasing OpenAI and Anthropic either; coverage names Moonshot AI's Kimi K3 as the direct rival in the open, low-cost model segment^[5], underscoring that the sharpest competition right now is happening among fast-moving open-weight players as much as against the closed labs. Independent testing has picked up on this quickly - one design-focused benchmarker found V4.1-Flash reaching 98% of a leading closed model's score at roughly 1.4% of the cost on everyday design tasks, framing it as evidence that open models are closing the gap with closed ones.

This isn't a parallel release that leaves the old model running. Starting at 04:00 UTC on September 14, 2026, every request to deepseek-v4-pro will be automatically routed to V4.1-Flash and billed at Flash-series rates, effectively retiring V4-Pro; the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names are also temporarily routed to the new model^[6]. DeepSeek says its own internal and external testing shows V4.1-Flash surpassing V4-Pro on performance, cost, speed and total completion time^[4], which is the stated justification for forcing the migration rather than offering it as an option. In practice, every V4-Pro integration gets switched to a materially different architecture within days of announcement, whether or not a given workload has been validated against it.

The context that makes this release notable isn't purely technical. In January 2025, DeepSeek's V3 and R1 release triggered a historic single-day selloff, with Nvidia's share price falling about 17% - roughly a $600 billion market-cap loss, the largest one-day drop for a single company in US stock market history at the time^[7]. V4.1-Flash's launch, matching or beating two current frontier closed models on coding benchmarks at a fraction of the price, produced no comparable market disruption^[7]. That gap is itself notable: eighteen months after the original shock, a Chinese lab shipping open, MIT-licensed weights that compete with top closed models has become an expected cadence rather than a surprise, following DeepSeek's pattern of splitting its main model line into separate Pro and Flash variants earlier in the V4 generation^[8].

The MIT-licensed weights are freely downloadable from Hugging Face with support for vLLM, SGLang, and Transformers, and no access restrictions gate who can use them^[1]. But openness doesn't mean accessibility: a 552B-backbone, 763B-total-parameter model still requires server-grade, multi-GPU hardware to self-host^[1], even with the KV-cache optimizations that make serving it at scale cheaper. The efficiency gains mostly benefit whoever is already running the infrastructure - DeepSeek's own API, or well-resourced enterprises - rather than putting a frontier-class model within reach of a single workstation.