DeepSeek V4.1 Flash Is Here — Now Available in ManusBot
AI-generated concept illustration representing asymmetric architecture, cache compression, and agent workflows; it is not a product screenshot.
On September 10, 2026, DeepSeek released DeepSeek V4.1 Flash. It is the smallest member of the company's new architecture family, designed to combine a higher capability ceiling, faster inference, greater throughput, and lower serving cost.
Most importantly for our users, DeepSeek V4.1 Flash is now integrated into ManusBot. There is no infrastructure to configure: open chat.manus.kim and use it for research, writing, coding, analysis, and everyday work.
What is new in this release?
According to DeepSeek's announcement and technical report, V4.1 Flash is a multimodal Mixture-of-Experts model with 552B total parameters and support for contexts of up to one million tokens. It natively accepts text and images. Its central idea is not simply adding parameters, but using compute and memory more efficiently.
For users, the release comes down to four improvements:
- stronger performance on agent-oriented tasks
- asymmetric input and output compute that reduces the cost of input-heavy work
- a much smaller KV cache for long conversations and complex workflows
- native vision-and-text understanding as the foundation for multimodal tasks
Causal Encoder-Decoder: compute where it matters
V4.1 Flash introduces a Causal Encoder-Decoder (CED) architecture. Its 40-layer Transformer is organized as a 20-layer causal encoder followed by a 20-layer decoder, so prompt prefill and output generation no longer use a fully symmetric computation path.
The model activates about 8B parameters per token during prefill and 16B during decoding. This is particularly useful for agent workflows that must read extensive context, call tools, and execute a sequence of steps: the model can consume rich input without spending the same amount of compute at every stage.
DeepSeek also reports a new pre-training approach and larger-scale reinforcement-learning post-training. In the company's Agentic Benchmark results, V4.1 Flash outperformed several frontier models, including DeepSeek V4 Pro. These are first-party evaluation results; real-world performance will still vary with the task, prompt, and tool environment.
A smaller KV cache lowers the cost of long agent tasks
Long conversations, large documents, and repeated tool calls can consume substantial high-bandwidth memory and storage through the KV cache. V4.1 Flash introduces several architectural changes to reduce that footprint.
DeepSeek reports a global KV cache size of 890 bytes per token:
- roughly one quarter of the HBM requirement of DeepSeek V4 Flash
- a persistent KV cache footprint of about one eighth of the previous generation
- approximately 437 times smaller than the first-generation DeepSeek model
These numbers are not only relevant to infrastructure teams. For agent products that depend on cache hits, long context, and multi-step tool use, a smaller cache creates room for higher throughput, lower cost, and smoother continuous-task experiences.
Native multimodality and a million-token context
V4.1 Flash jointly processes visual and textual information from pre-training rather than attaching vision to a text-only model as an afterthought. The official repository describes it as a native image-and-text input model that generates text.
Its maximum one-million-token context can accommodate much longer conversations, documents, and task traces. That makes the model a strong fit for research, code understanding, report synthesis, and multi-stage agent work that depends on substantial background material. Actual context limits and multimodal entry points depend on the capabilities exposed by each product.
Lower API pricing and a managed model transition
V4.1 Flash is available through the DeepSeek API under the model name deepseek-flash. DeepSeek also lowered pricing and retained time-of-day pricing, with off-peak rates set at half of peak-hour rates.
For compatibility, requests to deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash. DeepSeek plans to route deepseek-v4-pro requests to V4.1 Flash after 12:00 Beijing time on September 14, 2026, until V4.1 Pro becomes available, charging the V4.1 Flash rate.
Open weights, but infrastructure at a serious scale
DeepSeek published the V4.1 Flash weights, technical report, reference inference code, and encoding implementation on Hugging Face under the MIT License. This gives researchers and inference-framework teams an open starting point for adaptation and optimization.
Self-hosting a 552B-parameter model at scale still demands substantial GPU and storage capacity. DeepSeek's announcement invites organizations with roughly 2,000 GPUs and a storage cluster to discuss large-scale deployment. For most teams and individual users, accessing the model through a product or API is the practical route.
DeepSeek V4.1 Flash is live in ManusBot
The model's architecture is a natural fit for ManusBot: it is optimized for agent workflows and designed to use compute and cache efficiently across long inputs, multi-turn conversations, and complex tasks.
Try it in ManusBot to:
- summarize long documents and extract decisions or action items
- break down complex requirements into plans and structured outputs
- assist with coding, explain code, diagnose issues, and improve implementations
- organize research, compare information, and form clear conclusions
- draft articles, reports, emails, and product documents
No complex setup is required. Visit chat.manus.kim, enter a real task, and experience what DeepSeek V4.1 Flash can do inside ManusBot.