DeepSeek V4.1-Flash targets long-context agent costs with new cache architecture

The exact released model is DeepSeek-V4.1-Flash. Official access is through the DeepSeek API using deepseek-flash; the API docs list 1M context, 384K maximum output, JSON output, tool calls, Responses API, Anthropic-compatible API, chat-prefix completion, FIM completion in non-thinking mode, and vision support. Models & Pricing | DeepSeek API Docs

Pricing is materially lower than V4 Pro in the official table: for deepseek-flash, DeepSeek lists off-peak/peak cache-hit input at $0.003 / $0.006 per 1M tokens, cache-miss input at $0.15 / $0.30, and output at $0.60 / $1.20; V4 Pro is listed higher across those categories. Treat these as current posted prices, not guaranteed future economics, because the same page reserves the right to adjust prices. Models & Pricing | DeepSeek API Docs

Weights are available on Hugging Face, with the repository showing MIT licensing, minimal-inference references, prompt encoding tooling, and reproduction instructions for DeepSWE. The Hugging Face page also lists “model size” as 763B params, while the technical description distinguishes 552B backbone parameters plus additional components such as Engram memory; practitioners should avoid comparing a single headline parameter number without checking what is included. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

Architecture

DeepSeek’s technical report/model card describes V4.1-Flash as a multimodal sparse MoE with text and images jointly processed from the start of language-model pretraining. Its main architectural novelty is the Causal Encoder–Decoder / CED split: the model has 40 Transformer layers, organized as 20 encoder layers and 20 decoder layers. The stated inference goal is asymmetric compute: prompt/prefill tokens use fewer active parameters than generated/decode tokens. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

The KV-cache story is the core engineering contribution. Traditional decoder-only Transformers accumulate layer-specific key/value states for every attended token; long contexts therefore become memory- and bandwidth-heavy. DeepSeek says V4.1-Flash projects decoder global KV from the final encoder hidden state instead of maintaining independent global KV at every decoder layer. CSA2 then assigns layers to static attention modes — Full, Reindex, or Reuse — so some layers build KV/index structures, some borrow KV while recomputing sparse selections, and others reuse both. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

DeepSeek reports that FP4 main KV caching plus CSA2 reduces global KV cache footprint to 890 bytes per token, about one quarter of V4-Flash. The YouTube transcript’s “437× smaller than V1” claim is present in DeepSeek’s model-card figure caption as an official comparison across DeepSeek generations, but it should be treated as a vendor-reported internal baseline unless independently reproduced. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

A useful but explicitly caveated independent technical reading comes from an unlisted, AI-drafted ezyang study, which says it cross-checked the technical report, checkpoint config, reference inference code, and safetensor shapes. It derives the 890 bytes/token number from four layers holding a 512-wide FP4 main KV plus a 128-wide indexer key, with encoder caches storing one entry per two tokens and the decoder cache one entry per token. Because the page discloses that it awaits human editing, it is best used as an implementation-oriented interpretation, not a final peer-reviewed audit. An infra-oriented diagram of the DeepSeek-V4.1-Flash architecture

Other documented components include SWA Bounded Replay, intended to avoid persisting sliding-window KV to SSD; Single-Pass mHC, a residual-stream mixing method; Engram conditional memory, reported as sparsely accessed memory parameters; and DSpark speculative decoding. These are important for implementers because performance depends on kernels, cache layout, speculative acceptance, and memory placement, not just the abstract Transformer graph. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

Implementation and reproducibility

For self-hosting, the model is open-weight but not “easy local.” The official announcement even frames large-scale deployment as a conversation for organizations planning 2,000 GPUs plus a storage cluster, and vLLM Ascend’s deployment guide validates W8A8 colocated deployment on either two Atlas 800 A3 servers or four Atlas 800 A2 servers. That suggests the practical audience for full-precision or production self-hosting is labs and infrastructure teams, not individual desktop users. DeepSeek | Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.

The Hugging Face page provides vLLM and SGLang serving examples, recommends sampling parameters, and points to evaluation instructions for DeepSWE reproduction. However, a reproducible result still depends on exact prompt encoding, harness, reasoning effort, sampling, container environment, and tool permissions. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

V4.1-Flash extends DeepSeek’s earlier V4 direction: sparse MoE, long context, and KV-cache compression. DeepSeek’s V4 model card described V4-Flash as 285B parameters / 13B active per token, with earlier CSA/HCA cache compression. V4.1-Flash shifts to a larger 552B backbone but claims lower active compute on prefill and a smaller global KV footprint. Model properties

The broader industry context is that agentic systems are often bottlenecked by repeated long prompts, tool transcripts, repository context, browser state, and cache reuse. In such settings, KV cache is not an implementation detail; it becomes a cost driver. That is why a cache architecture can be commercially meaningful even if raw benchmark rankings remain contested.

Benchmark interpretation is the biggest contested area. AI Primer summarizes analyst concerns that V4.1-Flash’s strong results on older public Terminal-Bench versions and weaker results on newer versions could be consistent with public-benchmark contamination, while also noting that the allegation was tentative and that benchmark version changes alter tasks, resource allowances, instructions, environments, and verifiers. DeepSeek V4.1 Flash benchmark results draw analyst questions over possible contamination | AI Primer

Official benchmark tables use maximum reasoning effort and specific scaffolds. DeepSeek reports that code-agent benchmarks were run with DeepSeek Harness Minimal or benchmark-specific harnesses, 1M context, and temperature=1.0, top_p=0.95; scaffold choice changes results materially in its own table. This is useful transparency, but it also means buyers should benchmark with their own agent loop, tools, retry policy, and cost caps. deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

No independent safety report comparable to a full system card was found in the reviewed sources. Native vision, tool calls, long context, and high-output limits expand operational risk: prompt injection through images/documents, long-context data leakage, tool misuse, and runaway agent costs all need separate controls.

For business leaders, V4.1-Flash is most relevant where cost is dominated by long prompts and repeated cache hits: coding agents, repository maintenance, document-heavy workflows, security triage, research assistants, and browser/terminal automation. The official pricing makes cache hits extremely cheap relative to cache misses and output, so workflows that reuse stable context could see disproportionate savings. Models & Pricing | DeepSeek API Docs

For developers, the action item is to evaluate total task cost, not just per-token price. High reasoning effort may improve task completion but increase output/thinking tokens. Run A/B tests across reasoning effort levels, cache-hit ratios, scaffold choices, and failure-retry policies.

For infrastructure teams, the model is open-weight but specialized. Full benefit likely requires runtimes that understand CSA2, FP4 KV, SWA replay, Engram memory, and speculative decoding. Generic MoE serving may run the model but miss the economics that make it interesting.