DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, with changes to attention caching and the prompt format used to connect applications to the model. For operators evaluating the release, memory figures and integration requirements deserve separate checks.
The model card describes image and text inputs, text output and a context window of up to one million tokens. The repository and model weights carry an MIT licence.
What the cache figure measures
DeepSeek reports a global key-value, or KV, cache footprint of 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash’s. KV caching retains attention data that can be reused during generation.
The design combines a causal encoder-decoder structure with attention data shared across layers and FP4 caching. The decoder’s global cache is projected from the encoder’s final hidden states.
That global-cache measurement is not the model’s total serving-memory requirement. It does not establish that a deployment needs one-quarter of the hardware or delivers four times the throughput.
Prompt compatibility needs attention
The release has no Jinja-format chat template. Its encoding reference documents changes from V4, including revised tool-call tags, a numeric reasoning-effort setting and support for system messages within a conversation.
In that reference, reasoning effort accepts integers from 1 to 100. The setting applies in thinking mode and is inserted at the conversation’s beginning. Its completion parser expects well-formed output and does not repair malformed responses. Applications therefore need their own error handling.
For a team migrating an existing integration, those details make prompt and parser tests a practical first step before judging model quality.
The toolkit leaves deployment work to operators
DeepSeek’s separate deepseek-recipe toolkit provides Rust libraries and Python bindings for converting requests and parsing responses across several API formats. The caller still supplies inference, tool execution and HTTP transport.
Its documented gaps include enforcement of strict tool settings and JSON Schema output constraints, plus conversation retrieval through previous_response_id. Supporting a request format therefore does not establish complete behavioural compatibility.
The model’s inference example is explicitly a reference implementation, not a production serving engine. Deployment decisions still require workload-specific checks of memory, latency, output handling and failure behaviour.
