What Is DeepSeek V4.1 Flash? The 552B MoE, Explained

DeepSeek V4.1 Flash is the smallest model in DeepSeek's new architecture family, released on September 10, 2026. It is a multimodal Mixture-of-Experts (MoE) model with a 552B-parameter backbone, native image understanding, and a 1M-token context window. The MIT-licensed weights are on Hugging Face^[4].

It is not a minor refresh of V4 Flash. DeepSeek calls it the first member of a new architecture family designed for a higher capability ceiling, faster inference, higher throughput, and scaling to larger models — and it says that family's flagship, V4.1 Pro, is still to come^[1].

The headline trade is asymmetry: V4.1 Flash activates only 8B parameters per token while reading input (prefill) and 16B while generating output (decode). That split targets agent workloads, which spend most of their time reading fresh context, so cheaper reading makes cheaper agents^[2][5].

• 552B backbone parameters + 196B Engram conditional memory (763B total weights)
• Native multimodal vision (text + images) via a 32-layer DeepSeek-ViT
• 1M-token context; max output 384K tokens
• 8B active per token at prefill, 16B at decode
• MIT-licensed open weights on Hugging Face^[4]
• API model name changed to deepseek-flash^[1]

V4.1 Flash uses a Causal Encoder-Decoder (CED) design: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states^[4].

That projection is the trick behind the low activation count. Because the decoder reads from a shared encoded representation, the model does not need every parameter active on every layer. DeepSeek says CED effectively halves prefill computation for sufficiently long sequences, while single-token decode FLOPs rise only about 25% as context grows 256-fold from 4K to 1M tokens^[5].

On top of CED sit three more components. Compressed Sparse Attention 2 (CSA2) assigns each attention layer one of three static modes — Full, Reindex, or Reuse — to share main KV and indexer keys across layers. A Hierarchical Sparse Indexer restricts later indexing layers to a candidate pool built by the first Full-Mode layer, bounding deeper indexer cost independently of context length. Single-Pass mHC (manifold-constrained hyper-connections) revises residual-stream mixing^[4].

• 40 layers: 20-layer causal encoder + 20-layer decoder
• Hidden size 5120; 1 shared + 384 routed experts per MoE layer
• 6 routed experts activated per token
• CSA2 with Full / Reindex / Reuse layer modes
• Hierarchical Sparse Indexer: candidate pool of 2048 blocks of 8, best 512 kept per query
• Engram n-gram memory: 196B params across two hash tables (layers 1 and 14)
• DSpark speculative decoding head for faster generation^[4]

V4.1 Flash does not activate 8B parameters across the whole inference cycle. It activates 8B during prefill (reading input) and 16B during decode (generating output). This is the single most misreported spec of the release^[5].

The asymmetry is deliberate. An agent that repeatedly reads a repository, tool definitions, and conversation history is input-heavy: it ingests far more tokens than it emits. Making the read path cheap lowers the cost of exactly that pattern, which is why DeepSeek frames V4.1 Flash as an agent-first model.

It is also why the name is misleading. V4.1 Flash has a larger total footprint than the old V4 Flash (552B vs 284B backbone, a ~94% jump), and its decode path activates 16B versus V4 Flash's 13B. On the Hacker News launch thread, developers questioned whether a 552B model is 'not really flash anymore,' noting that self-hosting becomes far more demanding even if serving is more efficient^[5][3].

V4.1 Flash natively processes images and text, and generates text autoregressively. A vision encoder (DeepSeek-ViT, trained from scratch with 2D-RoPE and 3x3 pixel-unshuffle downsampling) and a two-layer MLP projector turn images into visual embeddings, processed jointly with text from the start of pre-training^[4].

This replaces the separate experimental V4-Flash-Vision-Exp model. Images enter the prompt as <|deepseek_image|> spans, and the router applies a separate routing bias for tokens inside an image span, so vision and text tokens do not compete for the same experts^[6].

• 32-layer ViT at hidden size 1024, patch 14
• 3x downsampling aligner; up to 1024 tokens per image (minimum 295,936 pixels)
• No limit on images per prompt
• MMMU-Pro 56.5, CVBench 77.9, DocVQA 95.6 on the base model^[4]

The engineering claim with the biggest cost impact is memory. V4.1 Flash stores its global KV cache at 890 bytes per token — roughly one-quarter of V4-Flash's HBM footprint and one-eighth of its SSD storage. Compared with DeepSeek V1, that is about a 437-fold reduction^[2][5].

The mechanism is a combination of CSA2, the hierarchical sparse indexer, and FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels). SWA Bounded Replay reconstructs missing sliding-window-attention states by replaying only the most recent n_win tokens instead of persisting them to SSD^[4].

DeepSeek says global KV stays in its persistent cache with a guaranteed lifetime of at least 72 hours, while short-lived SWA KV lives in a distributed memory pool provisioned from 10% of host DRAM. When the long-lived global cache survives but the SWA state does not, V4.1 Flash replays only the most recent window rather than full history^[5].

DeepSeek reports V4.1 Flash at max reasoning effort (100). On DeepSWE v1.1 it scores 74.2, ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0). On CyberGym it scores 88.1, ahead of every rival with a listed score^[3][4].

On Terminal-Bench 2.1 it reaches 90.6, the highest in DeepSeek's table. But the results are not uniformly dominant: Opus 5 leads on Terminal-Bench 3.0 (43.3 vs 30.0), Terminal-Bench 4.0 (51.8 vs 31.2), HLE (56.3 vs 36.8), and ProgramBench (37.0 vs 20.3). GPT-5.6 Sol leads GPQA Diamond and SEC-Bench Pro^[3][4].

A hidden cost variable sits inside those numbers: DeepSeek ran the table at max effort. Its own tests show effort 25 to 100 raises DeepSWE from 66.0 to 74.2 and Terminal-Bench 2.1 from 82.4 to 90.6, but consumes roughly 2.5x the output tokens. Effort levels between 60 and 80 recover most of the accuracy at less than half the token budget^[5].

The API model name is now deepseek-flash. Off-peak, one million cached input tokens cost $0.003, uncached input $0.15, and output $0.60. Peak rates are double. The new pricing took effect at 04:00 UTC on September 10, 2026^[3].

Legacy names still work: deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired but temporarily route to V4.1 Flash. The base URL is unchanged — https://api.deepseek.com for OpenAI-compatible calls, or https://api.deepseek.com/anthropic for the Anthropic format^[1][3].

Official partners WorkBuddy (including CodeBuddy) and OpenCode already support the model. Coding tools connect through documented flows: OpenCode's /connect deepseek command, Claude Code's ANTHROPIC_BASE_URL environment variables, and OpenClaw's onboarding^[2][7].

• model: deepseek-flash
• Off-peak: $0.003 cache hit / $0.15 cache miss / $0.60 output per 1M
• Peak (weekday mornings UTC): double those rates
• Concurrency limit: 2,500
• Supports Responses API, Anthropic API, JSON Output, Tool Calls, vision^[3]

V4.1 Flash replaces the entire old Flash line and is set to absorb V4-Pro traffic too. V4-Flash and V4-Flash-Vision-Exp are retired and routed to V4.1 Flash. From 04:00 UTC on September 14, 2026, every deepseek-v4-pro request routes to V4.1 Flash at Flash rates, until V4.1 Pro launches^[1][2].

DeepSeek says multiple parties' tests put V4.1 Flash ahead of V4-Pro on performance, cost, speed, and total runtime — its stated reason for phasing out the flagship. Because the old model ids are silently rerouted rather than deleted, teams with tightly tuned prompts or agent behavior should re-run regression tests rather than assume the swap is safe^[3][5].