V4.1 Flash
DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash) is DeepSeek's 552B-parameter sparse mixture-of-experts frontier-workhorse (released Sept 10, 2026), the first model built on the company's Causal Encoder-Decoder (CED) architecture with asymmetric activation of ~8B params on input and ~16B on output. Paired with a 1M-token (1,048,576) context window, native image modality, aggressive pricing and ~175 tps hosted throughput, it delivers near-flagship reasoning (GPQA Diamond 90.2, HLE-with-tools 63.9, AA Intelligence Index 39.5) and class-leading terminal/agentic coding (Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2) at a fraction of Pro-tier cost.
Empirical Evaluation Results
Architectural Profile & Capabilities
DeepSeek V4.1 Flash is the 2026 successor to the retired V4 Flash SKU and the debut product of DeepSeek's Causal Encoder-Decoder (CED) design. Unlike a symmetric dense or conventional MoE layout, CED deliberately decouples encoding from decoding compute: roughly 8B parameters activate during input/encoding and roughly 16B during autoregressive output generation, yielding a total pool of 552B parameters while keeping per-token inference cost close to a mid-size model. Against the October 2026 frontier — OpenAI GPT-6 Astra / GPT-5.6 Sol-Terra, Anthropic Claude Opus 5 / Sonnet 5, Google Gemini 3.8 Flash / Pro, DeepSeek V4 Pro / V4.1 Flash, xAI Grok 4.6, Qwen 3.8 and Moonshot Kimi K3 — V4.1 Flash positions as the high-throughput, low-cost reasoning tier rather than an absolute ceiling model. Capability profile: On the reasoning axis it is genuinely elite for a "Flash" tier, posting GPQA Diamond 90.2% (SOTA-flagged), HLE (no tools) 39.2% and HLE (with tools) 63.9% (SOTA-flagged), AA-LCR v1.1 84.0% (SOTA) and an AA Intelligence Index of 39.5 (SOTA-flagged). On the agentic/terminal axis it is exceptional: Terminal-Bench 2.1 at 90.6%, DeepSWE v1.1 at 74.2%, TAU-2 Airline 76.7% (SOTA) and AutomationBench 68.9% (SOTA). Mathematics is strong (MATH-500 97.3%, MathArena Apex 65.6%) but clearly effort-sensitive and version-sensitive — MathArena arXivMath drops from 57.6% (2026-06) to 43.9% (2026-08) and BrokenarXiv shifts 39.4% → 54.8% across the same window, indicating real benchmark drift on hard research-level math. Vision is supported natively (MMMU Pro 77.0%), with tool-augmented visual-agent tasks scoring far higher (BabyVision w/ tools 89.6%, Chartography w/ tools 78.9%). Cyber results are bifurcated: CyberGym 88.1% but ExploitGym only 15.3%, i.e. strong on defensive/knowledge tasks, weak on novel exploit synthesis. Knowledge reliability is a caveat zone — the AA-Omniscience Index is negative (-5.3), signaling calibrated abstention/hallucination risk on long-tail factual recall. Operational economics: pricing is set at $0.036/1M input and $0.6633/1M output (Benchmark Heaven record), with OpenRouter listing an even lower input rate; hosted output throughput is empirically ~175 tps (Benchmark Heaven), with manual telemetry spanning ~200–420 tps and cloud TTFT of ~0.9–1.0 s. The 1M-token context window makes it a viable substrate for whole-repository, whole-corpus and long-horizon agentic workflows, though effective recall beyond the multi-hundred-thousand-token range is expected to degrade relative to its strong mid-context AA-LCR score. Data-integrity note: DeepSeek's own and NVIDIA's model cards were found to report variant-configuration-divergent figures (e.g., GPQA Diamond values in the 89.9–90.9 band, AutomationBench appearing at both 68.9% and 54.8%, Terminal-Bench 4.0 at both 26.8% and 31.2%, and SWE-bench Multilingual at both 54.5% and 46.5%). Because these divergent numbers cannot be resolved to a single exact checkpoint/inference configuration, they have been EXCLUDED rather than added, per the anti-proxy exact-variant mandate. No sibling (V4 Pro, V4, V4 Flash, V3) scores were used.
552B total parameters with asymmetric activation: ~8B active on input/encoding and ~16B active on output/decoding — the defining feature of DeepSeek's first-generation Causal Encoder-Decoder (CED) architecture. Context window is 1,048,576 tokens (1M) with native image modality. Released September 10, 2026. Attention mechanism (MLA vs GQA) is NOT confirmed by official variant-specific documentation and is therefore not asserted. Test-time reasoning behavior is configurable (reasoning vs non-reasoning modes and effort levels are referenced by vendor and third-party trackers), so reported scores are effort-sensitive; benchmark configurations should be matched before comparison.
Recommended Workloads & Primary Use Cases
- •Effort-tier hedging: headline scores (90.6 Terminal-Bench 2.1, 74.2 DeepSWE) are at MAXIMUM reasoning effort; base/legacy configurations report far lower numbers (Terminal-bench 31.3, SWE-bench Verified 66.0), so unqualified benchmark comparisons are misleading — always report the effort tier
- •Throughput is deployment-dependent: optimized hosted endpoints hit ~200–420 tps, but local multi-GPU rigs drop to ~11 tps generation and TTFT can swing from <1s to multiple seconds under load; treat speed marketing numbers as best-case
- •Asymmetric 8B/16B activation split means prefill and decode behave differently — very long prompts are cheap to ingest but generation can be the bottleneck, and KV compression (FP4, sparse replay) risks subtle long-context recall degradation near the 1M-token boundary
- •Long-context quality decay: 1M-token window is nominal; RULER/MRCR-style needle tasks and cross-document reasoning typically degrade well before the ceiling, and precise intermediate-token recall is weaker than on short prompts
- •Vendor-reported benchmarks with limited independent replication: several scores come from DeepSeek's own model card/harness (and third-party aggregators flag harness looseness), so treat agentic-coding leadership claims as provisional pending neutral evaluation
- •Ecosystem and lifecycle churn: V4-Pro is being retired/rerouted to V4.1-Flash around Sept 14, 2026, so pinned model IDs and compatibility paths may shift; license reported as MIT but not independently verified from the license text
- •Effort/config sensitivity: reported benchmark values vary by reasoning mode and effort level (vendor cards show divergent GPQA Diamond and AutomationBench figures), so identical prompts can land on either the reasoning or non-reasoning scorecard — pin the checkpoint and effort setting for reproducibility.
- •Context degradation at the extreme: while the window is 1,048,576 tokens and AA-LCR v1.1 is strong at 84.0%, practical effective recall and multi-hop reasoning typically decay well before the full 1M boundary; chunk or retrieve for dense long-horizon tasks rather than assuming flat 1M performance.
- •Knowledge reliability gate: the AA-Omniscience Index is negative (-5.3), indicating elevated hallucination / weak-abstention risk on long-tail factual recall — require grounding, citation or retrieval for knowledge-critical enterprise outputs.
- •Latency profile: cloud TTFT of ~0.9–1.0 s plus reasoning-token generation means interactive chat and tight-loop UIs can feel deliberate; reserve it for throughput-oriented or asynchronous workloads, and note that throughput telemetry spans ~175 tps (measured) up to ~200–420 tps (hosted).
- •Pricing/spec divergence across aggregators: input/output token rates and multimodal framing differ between sources (e.g., Benchmark Heaven output price vs OpenRouter listing), and image modality is documented by some sources as native vision but not uniformly confirmed — validate the served SKU's modality and price before cost modeling.
- •Benchmark drift on hard math: MathArena arXivMath and BrokenarXiv moved materially between the 2026-06 and 2026-08 snapshots, so arithmetic on frontier-research math should be treated as version-dated rather than a stable capability claim.
Calls to V4.1 Flash are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by DeepSeek.
More Models from DeepSeek
Compare other engines available in this laboratory.
Terminal-Bench 2.1: 87.9%, DeepSWE: 62.7%, CyberGym: 83.3%, AutomationBench: 31.8%, Toolathlon-Verified: 74.1%, HLE w/ tools: 60.0%, GPQA Diamond: 90.1%, Artificial Analysis Intelligence Index: 53.2, AA Coding Index: 68.8, AA Agentic Index: 49.6 | The official GA release of V4 Pro, superseding the April preview (released Aug 13, 2026).
V4 Pro is a high-performance foundation model engineered by DeepSeek, accessible with Zero Data Retention on ARMES.
DeepSeek V4.1 Flash 1017 is the October 17, 2026 snapshot of DeepSeek V4.1 Flash — a 552B-parameter sparse Mixture-of-Experts model built on DeepSeek's Causal Encoder-Decoder (CED) architecture, activating only ~8B parameters during prefill and ~16B during decode. It pairs a 1,048,576-token context window with native image input, controllable reasoning effort, MIT-licensed open weights, and an extremely low price ($0.15/1M in, $0.60/1M out), positioning it as a cost-dominant agentic/coding workhorse.