DeepSeek-V4.1-Flash

An open-weight multimodal mixture of experts published on 10 September 2026 under an MIT licence. It runs 8 billion parameters per token while reading and 16 billion while writing, reads a million tokens of context, and stores its global key-value cache in 890 bytes per token, which is roughly a quarter of the previous Flash model. It leads DeepSeek's launch table on the long-horizon agent rows and sits behind Opus 5.0 and GPT-5.6 Sol on the hardest reasoning sets.

What it is

DeepSeek published DeepSeek-V4.1-Flash on 10 September 2026, with the weights on Hugging Face under an MIT licence and the model live on the company’s own API the same day. It is a multimodal mixture of experts: 552 billion backbone parameters, a context window of one million tokens, and a vision encoder trained from scratch, so images and text go in and text comes out.

The figure that governs the rest of the design is how little of the model runs at any one moment. V4.1-Flash activates 8 billion parameters per token while it reads a prompt and 16 billion while it writes an answer. DeepSeek calls the split asymmetric and aims it at agent workloads, where a long transcript is read many times over and the reply is short.

On the API it answers to deepseek-flash. Maximum output is 384,000 tokens, thinking mode is on by default, and the concurrency limit is 2,500.

The cache is the whole design

DeepSeek titled its technical report “Pushing the Limits of KV Cache Compression”, which says where the engineering went.

Every token a model has read leaves behind a key-value entry it consults while generating the next one. Across a million tokens of context that store becomes the dominant cost of serving the model, in memory while a request is live and in storage between turns. V4.1-Flash holds its global KV cache in 890 bytes per token. The model card puts that at roughly a quarter of DeepSeek-V4-Flash and about one 437th of DeepSeek-V1. The announcement gives operators the same result in the two units they buy: a quarter of the HBM and an eighth of the SSD storage of the previous generation.

It reaches the invoice because DeepSeek bills cached input separately and far more cheaply than fresh input. In the company’s words: “Cache-hit charges often account for a large share of agent costs. Compressing the cache cuts those costs significantly.”

How the architecture gets there

The backbone is a causal encoder-decoder, which DeepSeek abbreviates to CED: 40 transformer layers arranged 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, which is what allows the 8-billion prefill and 16-billion decode figures.

Compressed Sparse Attention 2 assigns every attention layer one of three static modes, Full, Reindex or Reuse, so layers share main KV and indexer keys and reuse top-K sparse-attention indices. Inside the decoder, a hierarchical sparse indexer restricts the later indexing layers to a candidate pool built by the first Full mode layer, which bounds the cost of deeper indexing however long the context runs.

Two more pieces do the compression itself. FP4 main KV caching stores the cache in E2M1 format with one E4M3 scale per 16 channels. SWA Bounded Replay reconstructs missing sliding-window states by replaying only the most recent n_win tokens, so those states never need writing to SSD, and the persistent cache lands at about an eighth of DeepSeek-V4-Flash.

Around that sit Single-Pass mHC, a revised residual-stream mixing scheme with a Mega-mHC kernel; Engram conditional memory, 196 billion parameters reached sparsely by token-based lookup and held alongside the 552-billion backbone; and DSpark speculative decoding, which drafts semi-autoregressively and verifies on a confidence schedule. Each mixture-of-experts layer carries one shared expert and 384 routed experts, and activates six routed experts per token.

Vision runs through DeepSeek-ViT, trained from scratch with 2D-RoPE and 3×3 pixel-unshuffle downsampling, and a two-layer MLP projector. Visual embeddings are processed jointly with text embeddings from the start of language-model pre-training rather than bolted on afterwards.

Pre-training used 45 trillion tokens of multimodal data, with sparse attention trained at a sequence length of 64,000 and the context extended to one million tokens at the 34-trillion-token mark. Post-training follows supervised fine-tuning, reinforcement learning and on-policy distillation, with DeepSeek stating that the changes are in the data pipeline rather than the algorithms. Reasoning effort is a controllable integer from 1 to 100.

DeepSeek V4.1 Flash specification card: 552B backbone, 196B Engram memory, 8B active at prefill and 16B at decode, a one million token context window, a 20-layer causal encoder feeding a 20-layer decoder for 40 layers in total, Compressed Sparse Attention 2, a hierarchical sparse indexer, an FP4 main KV cache at 890 bytes per token, DSpark speculative decoding, an MIT licence and the model string deepseek-flash. — The shape of the model, drawn from the DeepSeek model card and the release announcement of 10 September 2026.

What it scores

DeepSeek publishes a full instruct table, run at maximum reasoning effort with temperature=1.0 and top_p=0.95, against Opus 5.0, GPT-5.6 Sol, Kimi K3, GLM-5.3 and its own V4-Pro and V4-Flash.

V4.1-Flash leads that table on the long-horizon agent rows: Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2 resolved, CyberGym at 88.1, HLE with tools at 63.9, AutomationBench at 54.8 and Agent’s Last Exam at 31.8. On MathArena Apex it reaches 65.6, level with Kimi K3. Its Codeforces rating of 3471 tops that row too, which DeepSeek reports for its own models only: V4-Pro at 3348 and V4-Flash at 3289.

The frontier models hold the harder reasoning and long-form coding sets. Opus 5.0 takes HLE at 56.3 against 36.8, Terminal-Bench 3.0 at 43.3 against 30.0, Terminal-Bench 4.0 at 51.8 against 31.2, ProgramBench at 37.0 against 20.3 and NL2Repo-Bench at 75.3 against 64.0. On GPQA Diamond, V4.1-Flash records 90.9 behind GPT-5.6 Sol at 94.1 and Opus 5.0 at 93.4.

Read the method alongside the figures, because DeepSeek prints it. The code-agent rows run on DeepSeek’s own harness in Minimal mode with a one-million-token context; DeepSWE v1.1 uses the mini-SWE harness and SEC-Bench Pro the Claude Code harness, to meet each benchmark’s official setup; the visual agent rows use the Claude Code harness at 512,000 tokens.

The company then does something launch tables rarely do and publishes the spread across eight scaffolds. On Terminal-Bench 2.1 the same model records 90.6 on DeepSeek Harness Minimal, 90.3 on mini-SWE, 88.0 under Claude Code, 86.1 on Pi, 85.8 on DeepSeek Harness Standard and on PTC, 85.0 on OpenCode and 84.1 under Codex. On DeepSWE v1.1 it runs from 74.2 on mini-SWE down to 65.5 on OpenCode. The headline number in both cases is the best of eight, and DeepSeek shows the other seven.

One figure moves between DeepSeek’s own pages. The model card gives NL2Repo-Bench as 64.0; the change log entry for the same release gives 65.4.

The DeepSeek-V4.1-Flash model card on Hugging Face, showing the introduction and architecture section. — The model card on Hugging Face, captured 11 September 2026. Open the original.

What it costs

New API prices took effect at 04:00 UTC on 10 September 2026. Off-peak rates are half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, and every other hour is off-peak.

Per million tokens, deepseek-flash charges $0.003 off-peak and $0.006 peak for input that hits the cache, $0.15 and $0.3 for input that misses it, and $0.6 and $1.2 for output.

Set against deepseek-v4-pro at the same hour, which charges $0.022 and $0.044 cache-hit, $0.66 and $1.32 cache-miss, and $1.98 and $3.96 output, Flash output runs a little under a third of the flagship rate, its cache-miss input under a quarter, and its cached input roughly a seventh. The concurrency limits differ too: 2,500 against 500.

The DeepSeek models and pricing table listing deepseek-flash and deepseek-v4-pro with peak and off-peak rates per million tokens. — DeepSeek's live price table, captured 11 September 2026. Open the original.

Where the V4 Pro plan stands

The release announcement sets out a handover. It says third-party tests put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime, that DeepSeek is phasing V4-Pro out, and that “Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches.”

Two other pages from the same release record a different outcome. The change log entry dated 2026-09-10 and footnote 2 on the live price table both read: “In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged. We will provide further notice should there be any changes.”

The date both statements turn on is 14 September 2026. As of 11 September the price table lists deepseek-v4-pro as a current model on model version DeepSeek-V4-Pro-0813, with its own rates and its own concurrency limit, so the continuation is the position DeepSeek is currently serving.

The retirements that did complete are on the Flash side. V4-Flash and V4-Flash-Vision-Exp are retired, and the names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, routed to V4.1-Flash and billed at the Flash price. That answers the open question on our DeepSeek-V4-Flash-Vision-Exp page: the experimental vision model has a stable successor with weights, an MIT licence and multimodal support built into the backbone.

What to watch

V4.1-Pro. DeepSeek describes this model as the smallest in a new architecture family designed to scale to larger ones, and names V4.1-Pro as the release that ends the current V4-Pro routing arrangement.

Independent scaffold results. DeepSeek has published its own eight-scaffold spread, so the check available to everyone else is whether outside runs on Claude Code, Codex and OpenCode land near the 88.0, 84.1 and 85.0 the company reports.

The NL2Repo-Bench figure, which currently reads 64.0 on the model card and 65.4 in the change log.

Deployment paths. The company says it will work with the open-source community on inference support, and invites operators planning deployments of 2,000 GPUs and a storage cluster to make contact.