DeepSeek V4.1 Flash: New Architecture Cuts Agent Memory 4x
On September 10, 2026, DeepSeek released V4.1-Flash: a new architecture family, not a minor update. The model ships with a 1-million-token context window, a novel Causal Encoder-Decoder design, FP4 KV caching, and a 552-billion-parameter Mixture-of-Experts backbone. For enterprise teams running long-horizon AI agents, the headline number that matters is this: active high-bandwidth memory per token drops to roughly one-quarter of the previous V4-Flash generation.
This is the first entry in what DeepSeek is calling its V4.1 architecture lineage. A V4.1-Pro model is still ahead. For now, DeepSeek has routed its own V4-Pro API endpoint to Flash, which is the clearest signal the company can send that it considers Flash the stronger production choice today.
What Changed Architecturally
Four interlocking techniques drive the memory reduction:
Causal Encoder-Decoder (CED): The model splits its forward pass into an encoder phase (20 layers, 8B active parameters) for processing input and a decoder phase (40 layers, 16B active parameters) for generating output. During prefill, the model runs at roughly half the active parameter count of a traditional decoder-only architecture. For agents that spend more time reading context than writing tokens, this matters for throughput and cost.
Compressed Sparse Attention 2 (CSA2): Each attention layer is assigned one of three static modes: Full, Reindex, or Reuse. Subsequent layers can share the main KV cache and attention indices from earlier layers rather than computing and storing them independently. A Hierarchical Sparse Indexer further limits the candidate pool that deeper layers search, bounding cost at long context lengths.
FP4 KV Cache: The main global KV cache is stored in 4-bit floating-point precision (E2M1 format). DeepSeek applied quantization-aware training from the start rather than post-training compression, which avoids accuracy degradation. The local sliding-window attention cache, which is more numerically sensitive because it covers only the most recent tokens, stays at FP8.
SWA Bounded Replay: Instead of persisting the sliding-window attention cache on SSD between prefill and decode (as V4-Flash did), V4.1-Flash reconstructs it on demand from a temporary DRAM buffer, replaying the last 128 tokens. This eliminates the persistent SSD storage that occupied roughly half of V4-Flash’s storage footprint.
The combined result: a global KV cache of 890 bytes per token, versus approximately 3,560 bytes in V4-Flash. Active HBM usage drops 75%; persistent SSD usage drops to one-eighth. TechTimes reported that cache-hit charges routinely dominate inference bills for production agent workloads, so a 4x reduction is not a marginal improvement but a structural change in the economics.
The 1M-Token Context in Practice
The model was pre-trained on 45 trillion tokens, with sparse attention extended to 1 million tokens after 34 trillion tokens of training. Maximum output length is 384,000 tokens. Native multimodal support covers images and text jointly from the start of language-model pretraining, using a 32-layer vision encoder pre-trained on approximately 47 billion image-text pairs.
For enterprise builders, a 1-million-token context means:
- An entire software repository in a single session, no chunking required
- Multi-day agent conversations without context compaction hops
- Full financial or legal documents with embedded tables and images
- Long-running agentic loops that maintain state across hundreds of tool calls
This matters because context compaction, the process of summarizing or discarding old tokens to stay within a context limit, introduces both latency and accuracy risk. Agents that can hold more history without compacting make fewer errors on tasks that require reasoning across long time horizons.
Pricing and the V4-Pro Migration
DeepSeek published specific pricing with the launch:
| Tier | Cache-hit input | Cache-miss input | Output |
|---|---|---|---|
| Peak | $0.006 / M tokens | $0.30 / M tokens | $1.20 / M tokens |
| Off-peak | $0.003 / M tokens | $0.15 / M tokens | $0.60 / M tokens |
The V4-Pro endpoint migration is scheduled and mandatory. Starting at 04:00 UTC on September 14, 2026, all API calls to deepseek-v4-pro will route to V4.1-Flash at V4.1-Flash rates. V4-Flash and V4-Flash-Vision-Exp are already retired, with legacy endpoint names temporarily routing to V4.1-Flash during a transition window.
Teams on V4-Pro should note that output pricing drops from roughly $3.96 per million tokens (the rate set in August 2026 after a price increase) to $1.20 per million tokens at peak, or $0.60 off-peak. For high-volume agentic workloads, that is a significant change, but benchmarks should be re-run before the migration date given the architectural differences.
For context, this continues the pricing pressure tracked across the industry since late July 2026. As the enterprise AI pricing revolt coverage here explored, AI token costs have dropped faster than most enterprise budgets anticipated, and teams that locked in fixed infrastructure assumptions are now revisiting them.
What the Architecture Means for Enterprise AI Teams
The CED split has an underappreciated implication for multi-agent systems. When an orchestrator delegates a task to a subagent, the subagent typically reads a large context block (the task, tools, prior history) and generates a comparatively short output. The encoder-at-8B path handles the read; the decoder-at-16B handles the write. Workloads that are read-heavy relative to their output length, which describes most enterprise document processing, compliance review, and data extraction pipelines, benefit disproportionately.
The 196-billion-parameter sparse Engram memory component is separate from the main MoE backbone. DeepSeek has not published detailed documentation on it beyond the architecture overview, but early commentary on the OrcaRouter analysis suggests it acts as a persistent associative cache that the model can read from during long-context inference.
The reasoning effort setting (1-100, controllable via API) follows a pattern established by OpenAI and Anthropic: let builders trade reasoning depth for speed and cost depending on the task. A coding agent running background refactors might use effort=20; a compliance review agent reading a 400-page contract might use effort=80. This kind of tuning is increasingly standard in enterprise deployments, and V4.1-Flash’s controllable effort sits alongside similar controls in DeepSeek’s previous model generations.
Open Weights: Coming, With Infrastructure Requirements
DeepSeek has confirmed open weights are planned but has not given a public date. The practical constraint they published: approximately 2,000 GPUs plus a storage cluster. That puts V4.1-Flash in a different category from the smaller open-weight models that run on a single server. Enterprise teams planning on-premise or private-cloud deployment will need to size infrastructure accordingly.
For teams that need self-hosted inference today, V4-Flash’s weights remain available under a commercial license. V4.1-Flash weights will be worth waiting for given the 4x memory improvement, which directly reduces the GPU count required for a given context length.
What This Means for Enterprise AI Strategy
Three things to track before September 14:
-
Re-run your evaluations now. The V4-Pro to V4.1-Flash migration happens automatically. DeepSeek says V4.1-Flash outperforms V4-Pro, but internal evaluations on domain-specific tasks should confirm this before the routing change takes effect.
-
Reassess your context budget. If your agent workloads were designed around a 128K or 200K context limit, a 1-million-token model opens design options that were previously infeasible. The combination of larger context and lower per-token memory cost could eliminate entire middleware layers built to manage context compression.
-
Watch V4.1-Pro. DeepSeek describes V4.1-Flash as “the smallest model in the new architecture family.” V4.1-Pro, when it ships, will run the same CED and CSA2 architecture at a larger scale. The performance delta between Flash and Pro in prior generations was significant for complex reasoning tasks; expect the same pattern here.
For enterprise teams building long-horizon agents where KV cache costs dominate the inference bill, DeepSeek V4.1-Flash is the most architecturally significant open-weight release since Kimi K3. The 4x memory reduction is the kind of structural shift that changes what is economically feasible to build, not just what is technically possible.
To explore how this fits into your enterprise AI architecture, book a call with the Enera team.