Table of Contents
DeepSeek has released DeepSeek-V4.1-Flash, an open-weight multimodal Mixture-of-Experts model that is cheaper, faster, and stronger than its previous V4-Pro model for API use.
The release introduces a new Causal Encoder–Decoder architecture designed to reduce inference costs for long-context and agentic workloads. V4.1-Flash activates 8 billion parameters during input processing and 16 billion during output generation, while using a 552 billion-parameter MoE backbone.
The model is now live through the DeepSeek API under the model name deepseek-flash. DeepSeek also said requests to deepseek-v4-pro will be routed to V4.1-Flash starting at 04:00 UTC on Sept. 14, 2026, and billed at Flash prices until a future V4.1-Pro model is released.
DeepSeek Targets the Cost of Long-Running Agents
DeepSeek's main goal with the release is lowering one of the biggest costs behind long-running AI agents, the memory used for cached context.
Large language models use key-value, or KV, cache memory to avoid recomputing prior context as they generate new tokens. That cache becomes especially important for agents that keep long histories, inspect large codebases, use tools repeatedly, or work inside million-token contexts. In those settings, memory bandwidth and cache storage can become as important as raw compute.
DeepSeek says V4.1-Flash uses a Causal Encoder–Decoder design in which the decoder's global KV cache is projected from the encoder's final hidden states. The model handles input and output asymmetrically. Input-heavy prefill uses fewer active parameters, while output generation uses more.
The company pairs that design with Compressed Sparse Attention 2, FP4 KV cache, and a system it calls SWA Bounded Replay, which reconstructs sliding-window KV cache instead of storing it on SSD. The result is a global KV cache of 890 bytes per token, about one-quarter the HBM footprint of V4-Flash, about one-eighth the persistent or SSD cache requirement, and roughly 437 times smaller than DeepSeek-V1's KV cache.
For developers, this means lower costs for agent workloads that read far more context than they write. That includes coding agents, research agents, workflow automation systems, and multimodal assistants that may need to carry large documents, screenshots, tool outputs, or repository context through a task.
What V4.1-Flash Supports
DeepSeek-V4.1-Flash is a native multimodal model that accepts images and text and generates text. The hosted API supports thinking and non-thinking modes, tool calls, JSON output, the Responses API, Anthropic-compatible API access, chat prefix completion in beta, fill-in-the-middle completion in non-thinking mode, and native vision.
The release also retires the earlier V4-Flash and V4-Flash-Vision-Exp models. DeepSeek says the compatibility aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp will still be accepted for now, but requests using those names are served by V4.1-Flash and billed at Flash prices.
The model supports a 1 million-token context window and a maximum output length of 384,000 tokens; DeepSeek trained the model from scratch on 45 trillion multimodal tokens, extended sparse attention to 1 million tokens during pretraining, and used supervised fine-tuning, reinforcement learning, and on-policy distillation during post-training.
The model also features a 196 billion-parameter sparse Engram memory, a vision encoder trained from scratch, and a controllable reasoning-effort setting from 1 to 100. V4.1-Flash is the smallest model in DeepSeek's new architecture family, with V4.1-Pro planned later but not yet released.
API Prices Fall Below V4-Pro
DeepSeek's new hosted pricing makes the cost reduction explicit. For V4.1-Flash, the company lists peak prices of $0.006 per 1 million cache-hit input tokens, $0.30 per 1 million cache-miss input tokens, and $1.20 per 1 million output tokens. Off-peak prices are half that: $0.003, $0.15, and $0.60.
By comparison, V4-Pro pricing before the Sept. 14 reroute is $0.044 per 1 million cache-hit input tokens, $1.32 per 1 million cache-miss input tokens, and $3.96 per 1 million output tokens during peak hours, with off-peak prices again half of peak.
Peak hours run from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday. Weekends and public holidays are off-peak.
The new model also raises the hosted concurrency cap to 2,500, compared with 500 for V4-Pro. DeepSeek said WorkBuddy, CodeBuddy, and OpenCode already support V4.1-Flash, and the consumer DeepSeek app has been updated with the model in Fast, Expert, and Visual modes.
Company Benchmarks Emphasize Coding and Agent Tasks
DeepSeek says V4.1-Flash surpasses V4-Pro across performance, cost, speed, and total runtime after testing by multiple parties. Its published benchmark tables show the new model performing especially well on coding, software-engineering, terminal, and agent benchmarks.
The company-reported figures put V4.1-Flash ahead of V4-Pro on Codeforces rating, Terminal-Bench 2.1, Terminal-Bench 3.0, Terminal-Bench 4.0, DeepSWE v1.1, CyberGym, SEC-Bench Pro, ExploitGym, HLE with tools, Automation-Bench, and Agents' Last Exam. DeepSeek's materials also show it close to or ahead of several closed models on some agent-focused tests.
Those numbers should be read as company-published results, not independent third-party verification. The same charts show V4.1-Flash trailing some closed models on certain knowledge, vision, and security evaluations, including portions of HLE, ProgramBench, NL2Repo-Bench, Chartography, BabyVision, and ZeroBench.
That mix is still consistent with DeepSeek's positioning. V4.1-Flash is being framed less as a universal maximum-capability model and more as an efficient, open-weight model for developers running agents, code workflows, tool use, and long-context tasks at lower cost.
The release follows DeepSeek's July V4-Flash update, which also focused heavily on agent and coding benchmarks.
Bart Hillerich
Open Weights Keep Pressure on Hosted AI Models
DeepSeek's decision to release weights under an MIT license keeps the model in the open-weight ecosystem while also using the hosted API as the easiest route for most developers. The Hugging Face repository includes the model card, technical report, and files for self-hosting, although the scale remains far beyond casual local deployment.
The company said open-source inference support is coming and invited teams planning deployments around roughly 2,000 GPUs plus a storage cluster to contact it. That detail underlines the practical split in open-weight AI. Developers can inspect, adapt, and deploy the model under a permissive license, but production-scale serving still requires substantial infrastructure.
V4.1-Flash also adds pressure on closed-model providers competing for coding and agent workloads. Its combination of open weights, native vision, million-token context, sharply lower API pricing, and V4-Pro API replacement gives developers a clearer reason to test DeepSeek's stack against more expensive hosted models.
The release also makes DeepSeek's roadmap more explicit.
DeepSeek presents V4.1-Flash as the first smaller model in the architecture family rather than its endpoint, with V4.1-Pro still ahead. For now, DeepSeek is using Flash to replace its own V4-Pro API path and to argue that efficient architecture, not just larger parameter counts, will define the next stage of agent infrastructure.
