---
type: "post"
id: "the-technical-report-describes-a-40-layer-causal-ba1f4b"
slug: "the-technical-report-describes-a-40-layer-causal-ba1f4b"
title: "The technical report describes a 40-layer Causal Encoder-Decoder network divided evenly between encoder and decoder layers."
description: "Thesis: DeepSeek-V4.1-Flash treats KV-cache footprint and movement as central constraints on long-horizon agent serving, not merely as secondary implementation details."
retrieval_nugget: "Thesis: DeepSeek-V4.1-Flash treats KV-cache footprint and movement as central constraints on long-horizon agent serving, not merely as secondary implementation details."
published_at: "2026-10-17"
updated_at: "2026-10-17"
record_date: "2026-10-17"
date_kind: "published_at"
topics: ["inference","agent-runtime","long-running-agents"]
source_urls: ["https://arxiv.org/pdf/2609.19969"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:the-technical-report-describes-a-40-layer-causal-ba1f4b"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:78dbd04f4cfbf3bdc394febf8d5aee000538d02cf12d42170b52713a086a1665","reviewed_at":"2026-10-17T11:01:58.302Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"DeepSeek-V4.1-Flash treats KV-cache footprint and movement as central constraints on long-horizon agent serving, not merely as secondary implementation details.","observed_facts":[{"text":"DeepSeek-AI describes a 552B-parameter multimodal MoE model with context up to one million tokens and a 40-layer Causal Encoder-Decoder split evenly between encoder and decoder layers.","source_urls":["https://arxiv.org/pdf/2609.19969"]},{"text":"The report says global KV cache is 890 bytes per token, about one quarter of DeepSeek-V4-Flash, and persistent KV footprint is roughly one eighth with SWA Bounded Replay.","source_urls":["https://arxiv.org/pdf/2609.19969"]}],"mechanism":"Cross-layer KV reuse in Compressed Sparse Attention 2 and FP4 KV caching reduce the global cache, while SWA Bounded Replay reduces persistent cache requirements. The architecture also reports 8B active parameters per token at prefill versus 16B at decode, separating the cost profile of prompt ingestion from generation.","why_now":"DeepSeek frames long-horizon agent workloads as increasingly input-heavy, making prefill compute, memory capacity, and cache transfer practical deployment bottlenecks. If these reported reductions hold under real workloads, they could widen the feasible operating envelope for extended agent sessions.","implications":["The first implication is that deployment comparisons for long-context agents should measure cache bytes, persistence, and transfer costs alongside benchmark quality and token pricing.","Reported benchmark scores may attract attention, but independent replication is needed before treating them as comparable evidence of capability."],"evidence_boundary":"The supplied primary source is DeepSeek-AI’s technical report; its claims and benchmark results are not independently validated by the supplied evidence, and it does not establish production performance or cost.","watch_conditions":["The first watch condition is whether released checkpoints and independent evaluations reproduce the cache-footprint and benchmark claims under comparable long-horizon workloads.","Also watch actual serving throughput, latency, and hardware requirements rather than relying on cache size alone."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied for this specific model report, so this begins a distinct record focused on its reported architecture and cache strategy."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/the-technical-report-describes-a-40-layer-causal-ba1f4b/","markdown":"https://newruntime.com/posts/the-technical-report-describes-a-40-layer-causal-ba1f4b.md","json":"https://newruntime.com/posts/the-technical-report-describes-a-40-layer-causal-ba1f4b.json"}
---

# The technical report describes a 40-layer Causal Encoder-Decoder network divided evenly between encoder and decoder layers.

## Retrieval answer

Thesis: DeepSeek-V4.1-Flash treats KV-cache footprint and movement as central constraints on long-horizon agent serving, not merely as secondary implementation details.

Thesis: DeepSeek-V4.1-Flash treats KV-cache footprint and movement as central constraints on long-horizon agent serving, not merely as secondary implementation details.

Observed facts: DeepSeek-AI describes a 552B-parameter multimodal MoE model with context up to one million tokens and a 40-layer Causal Encoder-Decoder split evenly between encoder and decoder layers. The report says global KV cache is 890 bytes per token, about one quarter of DeepSeek-V4-Flash, and persistent KV footprint is roughly one eighth with SWA Bounded Replay.

Mechanism: Cross-layer KV reuse in Compressed Sparse Attention 2 and FP4 KV caching reduce the global cache, while SWA Bounded Replay reduces persistent cache requirements. The architecture also reports 8B active parameters per token at prefill versus 16B at decode, separating the cost profile of prompt ingestion from generation.

Why now: DeepSeek frames long-horizon agent workloads as increasingly input-heavy, making prefill compute, memory capacity, and cache transfer practical deployment bottlenecks. If these reported reductions hold under real workloads, they could widen the feasible operating envelope for extended agent sessions.

Implications: The first implication is that deployment comparisons for long-context agents should measure cache bytes, persistence, and transfer costs alongside benchmark quality and token pricing. The report’s benchmark table includes 74.2 on DeepSWE v1.1 and 54.8 on Automation-Bench, but these figures are author-reported.

Evidence boundary: The supplied primary source is DeepSeek-AI’s technical report; its claims and benchmark results are not independently validated by the supplied evidence, and it does not establish production performance or cost.

Watch conditions: The first watch condition is whether released checkpoints and independent evaluations reproduce the cache-footprint and benchmark claims under comparable long-horizon workloads. Also watch actual serving throughput, latency, and hardware requirements rather than relying on cache size alone.

New branch reason: No prior New Runtime record is supplied for this specific model report, so this begins a distinct record focused on its reported architecture and cache strategy.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- topic hub: [Agent Runtime](https://newruntime.com/topics/agent-runtime.md): Explore the agent-runtime topic hub.
- topic hub: [Long Running Agents](https://newruntime.com/topics/long-running-agents.md): Explore the long-running-agents topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)