DeepSeek released V4.1-Flash on September 10, 2026 at 04:00 UTC, a 552-billion-parameter multimodal model engineered specifically to relieve the memory bottleneck that has come to dominate production AI agent deployments. The release went live alongside a 50-page technical report on Hugging Face and debuts DeepSeek’s V4.1 model series as a distinct architecture lineage, not an incremental update to V4. The headline figure: the active key-value cache memory required for long-running agents is one-quarter of what the prior V4-Flash generation needed, while persistent SSD storage for that cache drops to approximately one-eighth.
For developers running production agent workloads, tools that spend hours browsing the web, executing code, reading error logs, and accumulating hundreds of thousands of tokens of context, the announcement carries a direct cost consequence. Cache hit charges on agentic workloads routinely account for the majority of inference spending, which means a model that needs one-quarter the high-bandwidth memory and one-eighth the SSD storage is a structural economic change rather than a marginal optimization. DeepSeek has made the transition mandatory: starting September 14, 2026 at 04:00 UTC, all API traffic directed to the deepseek-v4-pro endpoint will be automatically rerouted to DeepSeek V4.1-Flash at V4.1-Flash rates.
DeepSeek V4.1-Flash Targets the Memory Wall That Outgrew Model Weights
Every time a large language model processes a prompt, it generates and stores intermediate mathematical representations known as key-value pairs for each token it has seen, keeping them in a KV cache so it does not have to recompute them on subsequent passes. For a basic chatbot this overhead is manageable. For a long-running agent that spends hours reading tool outputs, inspecting code, and iterating on plans, the cache inflates into a substantial memory burden, and at sufficient context length the KV cache outgrows the model weights themselves. The model’s “thinking notes” effectively become larger than the model. For enterprise teams running agents at scale, this makes cache management the binding operational constraint, not raw model performance.
DeepSeek’s engineering logic in building DeepSeek V4.1-Flash was to attack that constraint directly, rather than continuing to optimize the attention computation the way V4 did. The result is a global KV cache of 890 bytes per token, roughly one-quarter of V4-Flash and approximately 1/437th of DeepSeek-V1’s per-token global KV cache size from two years ago. The reduction comes from four interlocking architectural techniques.
How DeepSeek V4.1-Flash Achieves Its Four-Way Memory Reduction
The first technique is a Causal Encoder-Decoder split. DeepSeek V4.1-Flash reorganizes its 40 Transformer layers into two asymmetric halves: a 20-layer causal encoder and a 20-layer decoder. During the prefill phase, when the model reads a long input such as hundreds of pages of accumulated tool outputs, only the 20 encoder layers handle the full context. The decoder’s global KV cache is then synthesized directly from the encoder’s final representations rather than being recomputed independently by each decoder layer. The model activates only 8 billion parameters per token during input processing compared to 16 billion during text generation. The design was inspired by YOCO, a 2024 research approach demonstrating that global KV caches could be shared across decoder layers.
The second technique is Compressed Sparse Attention 2 (CSA2). Where V4 used a static attention compression scheme, DeepSeek V4.1-Flash assigns every Transformer layer one of three operating modes: Full, which computes a fresh cache and selects the most relevant positions; Reindex, which shares an existing cache with other layers but independently chooses which entries to attend to; and Reuse, which shares both the cache and the prior layer’s attention selections outright. Deeper layers borrow the work of shallower ones, eliminating redundant KV storage across the network’s depth.
The third technique is FP4 cache quantization. DeepSeek V4.1-Flash stores its main global KV cache in FP4 precision, a 4-bit floating-point format, rather than the FP8 used by V4. The move from 8-bit to 4-bit roughly halves memory per cache entry. DeepSeek applied quantization-aware training, so the model was trained from the start to work with FP4 representations, preventing the accuracy degradation that comes from compression bolted on after the fact. The local sliding-window attention cache, which covers only the most recent tokens and is more numerically sensitive, remains stored at FP8.
The fourth technique eliminates persistent SWA storage entirely. Global KV cache entries can remain useful for days, representing accumulated context an agent might reference at any point. Sliding-window attention state typically matters for only the most recent few minutes of an active session, covering a bounded local window of 128 tokens in DeepSeek V4.1-Flash’s design, and becomes irrelevant quickly. In V4’s prior architecture, SWA KV occupied nearly half of persistent cache capacity on SSD drives, sitting there for over 72 hours unnecessarily. DeepSeek V4.1-Flash stops persisting SWA state to SSD. Instead, a temporary distributed memory pool drawn from 10% of each server machine’s host DRAM holds SWA state for the brief window it is useful. A technique called SWA Bounded Replay reconstructs that state after a session pause or preemption by replaying only the last 128 tokens rather than performing an expensive full reconstruction.
DeepSeek V4.1-Flash Activates 8 to 16 Billion Parameters Despite 552B Total
DeepSeek V4.1-Flash uses Mixture-of-Experts routing to keep computation lean despite its 552-billion-parameter total. Each Transformer block contains one shared expert and 384 routed experts, but only six routed experts activate for any given token, an activation rate of roughly 1.6%. Combined with the CED architecture, the model activates just 8 billion parameters per token during input prefill and 16 billion during text generation.
The V4.1 architecture also integrates a 196-billion-parameter conditional memory module called Engram, distinct from the external startup Engram that raised $98M earlier this year with a similar concept. DeepSeek’s Engram uses n-gram lookup tables of roughly 16 million entries each, specialized in memorizing patterns so the Transformer backbone can focus its expensive computation on reasoning rather than rote recall. The lookup tables reside in host memory and are prefetched during inference to avoid becoming a bottleneck. The full model trained on 45 trillion tokens, with the context window extended from 64K to 1 million tokens after 34 trillion training tokens had been processed. Native multimodal support, with text and images processed jointly from the start of language-model pretraining, arrives through a 32-layer vision encoder pre-trained on approximately 47 billion image-text pairs.
Among the most deployable features is a continuously adjustable reasoning effort setting, an integer from 1 to 100, which adjusts how strongly the training reward penalized extra reasoning tokens rather than imposing a hard cap, giving developers direct control over the latency-versus-quality tradeoff on each request. With its combination of mandatory endpoint migration, structural memory reductions, and granular reasoning control, DeepSeek V4.1-Flash positions the V4.1 lineage as an architecture built for the agent era rather than for chatbot-era benchmarks.
Source: Source article