DeepSeek-V4.1-Flash: 552B MoE at 890 Bytes per Token

DeepSeek released DeepSeek-V4.1-Flash on September 14, 2026 to attention for its benchmark results, but the architectural specifics are the more consequential story. The model is a 552B-parameter Mixture-of-Experts system that holds only 8B parameters active during prefill and 16B during decoding, compresses its global KV cache to 890 bytes per token, supports a 1-million-token context window, and ships under an MIT license with reference inference code. For practitioners running long-context agents at scale, those numbers determine whether deployment is economically viable — and they are the result of deliberate co-design across multiple architectural layers, not parameter scaling.

The problem DeepSeek is addressing is structural. Agentic workloads are input-heavy: a coding agent reading a large repository and conversation history may process hundreds of thousands of tokens before emitting a few thousand tokens of output. Traditional decoder-only Transformers treat every token identically regardless of whether it is being ingested or generated. As production AI deployments repeatedly demonstrate, architecture rather than raw model intelligence is what breaks under operational load, and symmetric prefill/decode costs are a compounding tax on agentic pipelines.

Causal Encoder-Decoder: Splitting the Compute Budget

V4.1-Flash addresses the prefill/decode imbalance with a Causal Encoder-Decoder (CED) architecture comprising a 20-layer causal encoder followed by a 20-layer decoder. During prefill, the decoder layers obtain their global KV representations from the encoder's final output rather than computing independent global KV states layer by layer. This consolidation drives active parameters down to 8B during prefill while allowing 16B to remain active during decode, where token-by-token reasoning demands more compute. The result is a compute profile deliberately calibrated for agent workloads: cheap ingestion, richer generation — which is precisely where architectural specificity starts outperforming raw GPU scaling.

CSA2 and FP4: Compressing the KV Cache

At 890 bytes per token, one million tokens of global KV data occupies approximately 890 MB — and DeepSeek reports this represents roughly one-quarter of the HBM demand of V4-Flash for its global KV cache. Persistent KV storage is further cut to around one-eighth of V4-Flash through SWA Bounded Replay, which discards some recent attention states and reconstructs a small local window on demand.

The mechanism behind the cache reduction is Compressed Sparse Attention 2 (CSA2), which eliminates redundant per-layer KV work through three operational modes:

A hierarchical sparse indexer further limits per-layer search cost: an upstream stage narrows a million-token context down to a smaller candidate set, and downstream layers search only within that subset. FP4 KV caching then compresses stored representations into a lower-precision format. The compound effect of cross-layer KV sharing, sparse hierarchical indexing, and FP4 storage is what lands at the 890-byte-per-token figure.

Two additional mechanisms reduce memory traffic during generation. Single-Pass mHC reorganizes activation mixing so repeated reads of the same data are consolidated, cutting memory traffic for that operation by roughly 50%. DSpark speculative decoding generates draft tokens that the main model verifies in batch, allowing multiple tokens to be accepted per forward pass and increasing effective throughput.

Benchmark Results on Agent Tasks

V4.1-Flash improves on both V4-Flash and V4-Pro across the agent-oriented evaluation suite: 74.2 on DeepSWE v1.1 (versus 54.4 and 62.7), 90.6 on Terminal-Bench 2.1 (versus 82.7 and 87.9), 88.1 on CyberGym (versus 76.7 and 83.3), 54.8 on AutomationBench (versus 37.7 and 43.2), and 31.8 on Agent's Last Exam (versus 25.2 and 25.7). The model was trained from scratch on 45 trillion multimodal tokens, accepts text and image input, and pairs its 552B backbone with a 196B-parameter Engram conditional-memory component accessed sparsely rather than activated uniformly across tokens.

AI Mastery Analysis

The architectural stack in V4.1-Flash is a coherent systems-engineering answer to a specific deployment problem rather than a general-purpose scaling exercise. Three implementation consequences stand out.

First, the 8B/16B active-parameter asymmetry only delivers its theoretical cost advantage if the serving infrastructure can route prefill and decode to separate compute pools or batch them differently — a non-trivial operational requirement that teams running local AI infrastructure stacks will need to engineer explicitly. Second, CSA2's three modes introduce a layer-assignment hyperparameter: the ratio of Full, Reindex, and Reuse layers is a design choice whose sensitivity to task distribution is not fully characterised in public documentation, creating a tuning surface downstream fine-tuners will encounter. Third, the Engram 196B conditional-memory component adds to total model storage and loading time; teams evaluating the model on active-parameter count alone may underestimate its hardware footprint at rest.

The source author's own testing, cross-referenced against Artificial Analysis and OpenRouter data, finds that models such as GLM-5.3-Flash currently offer stronger raw intelligence at lower cost. V4.1-Flash's primary practical advantage is output speed, which is consistent with its architectural priorities. The release is therefore more valuable as a methods contribution to the open-weight ecosystem than as a deployment recommendation for intelligence-per-dollar optimisation today.

The MIT license and published reference inference code matter beyond legal permissiveness: techniques like asymmetric encoder-decoder prefill, FP4 KV caching, and CSA2-style cross-layer KV sharing are now available for integration into open inference runtimes. Systems-engineering gains across the full inference stack are compounding in ways that rival headline parameter scaling, and whether V4.1-Flash's specific techniques propagate through open inference frameworks will be the real measure of this release's long-term impact.

Primary source

Why DeepSeek-V4.1-Flash Is Such an Exciting Open Model Release — KDnuggets