V4.1 Flash
DeepSeek-V4.1-Flash is DeepSeek's first Causal Encoder-Decoder (CED) frontier model: a 552B-parameter multimodal sparse MoE that activates only 8B parameters during prefill and 16B during decode, with a 1M-token context window, native image understanding, and aggressive FP4/compressed KV caching. It tops agentic terminal and software-engineering boards (90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1 at max effort) at roughly $0.15/$0.60 per 1M tokens, and supersedes V4-Pro in production traffic.
Empirical Evaluation Results
Architectural Profile & Capabilities
DeepSeek-V4.1-Flash was released 2026-09-10 and is the smallest — and first — model in DeepSeek's new CED family, published open-weight with a companion technical report (DeepSeek_V41_Tech_Report.pdf) on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash). It is served as `deepseek-flash` on the DeepSeek API with native multimodal support, and it is the successor routed to by V4-Pro, which retires around 2026-09-14.\n\nARCHITECTURE: The model is a multimodal Mixture-of-Experts (MoE) with a 552B-parameter backbone (plus a reported ~196B of Engram parameters) and a 1M-token (1,048,576) context window with a max output of ~384K. The defining innovation is the Causal Encoder-Decoder (CED): a 40-layer Transformer split into a 20-layer causal encoder and a 20-layer decoder, where the decoder's global KV cache is projected from the encoder's final hidden states. MoE routing uses 1 shared expert plus 384 routed experts with a top-6 sparse gate. Crucially, the activation profile is asymmetric — 8B active parameters for input (prefill) tokens and 16B for output (decode) tokens — which lowers prefill cost while giving the decoder a richer subnet for generation quality. Attention relies on compressed KV caching (FP4 main KV, Compressed Sparse Attention 2, SWA Bounded Replay, ~890 bytes/token global KV footprint) rather than plain GQA; DeepSeek reports roughly a 4x KV-cache memory reduction versus the previous Flash generation. Vision is handled by DeepSeek-ViT, trained jointly with the language model rather than bolted on. Generation is autoregressive — there is no diffusion or non-autoregressive decoding in the public record.\n\nCAPABILITY & POSITIONING: DeepSeek and OpenRouter position V4.1-Flash as outperforming V4-Pro on performance, speed, and task-completion time while being far cheaper. At maximum reasoning effort the official scorecard reports 74.2 on DeepSWE v1.1 and 90.6 on Terminal-Bench 2.1 — numbers that outscore Claude Opus 5 and OpenAI GPT-5.6 Sol on those agentic-coding suites. On the DeepSeek API changelog, base SWE-bench Verified is 66.0, SWE-bench Multilingual 54.5, and legacy Terminal-bench 31.3 (a non-reasoning-effort configuration). MATH-500 reaches 97.3% pass@1. Independent aggregators (Artificial Analysis, BenchLM, LLM-Stats) track V4.1-Flash as a top cost-efficiency performer in its tier.\n\nSERVING & SPEED: The CED design plus FP4 KV caching targets faster inference and higher throughput. Real-world reports put output speed at roughly 200–420 tokens per second on optimized hosted endpoints (some isolated claims above 500), with Artificial Analysis' fastest tracked DeepSeek model near 195 tps and cloud-API TTFT commonly around 0.9–1s. Local multi-GPU rigs are far slower (~11 tps generation), so throughput is strongly deployment-dependent. Pricing on OpenRouter is $0.15/1M input and $0.60/1M output, with DeepSeek's off-peak cached-input rate reported as low as $0.003/1M. Weights are open (reported MIT) with day-one vLLM support on NVIDIA and AMD.\n\nBOTTOM LINE: V4.1-Flash is a deliberately lopsided efficiency machine — enormous capacity, tiny active compute, 1M context, cheap tokens — purpose-built for long-horizon agents, terminal/computer-use loops, and high-throughput multimodal pipelines rather than for raw single-shot reasoning depth against a full-size Opus-5-class model.
552B-parameter MoE backbone (~196B additional Engram parameters reported), 40 Transformer layers split as a 20-layer causal encoder + 20-layer decoder, where the decoder's global KV cache is projected from the encoder's final hidden states. Sparse routing: 1 shared expert + 384 routed experts, top-6 active per token. Asymmetric activation: ~8B parameters active during prefill/input and ~16B during decode/output, keeping per-token compute low relative to total size. Attention uses compressed/sparse KV caching (FP4 main KV, Compressed Sparse Attention 2, SWA Bounded Replay; ~890 bytes/token KV footprint) rather than plain GQA, reducing KV memory ~4x vs the prior Flash generation. Native vision via DeepSeek-ViT, jointly pre-trained with the language model. Autoregressive generation; max output ~384K tokens; 1,048,576-token context. Hybrid reasoning with selectable effort tiers (low/high/max) and controllable thinking effort; not a diffusion or non-autoregressive model.
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
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.
Terminal-Bench 2.1: 82.7%, CyberGym: 76.7%, Toolathlon-Verified: 70.3%, DSBench-FullStack: 68.7%, DSBench-Hard: 59.6%, DeepSWE: 54.4%, NL2Repo: 54.2%, Agents' Last Exam: 25.2%, AutomationBench Public: 25.1%, Artificial Analysis Intelligence Index: 50 | The official GA release of V4 Flash, superseding the April preview (released Jul 31, 2026).