DeepSeek-V4.1-Flash is an open multimodal MoE model with 552B total parameters (+196.6B Engram parameters and 14.2B DSpark MTP parameters) and a context window of up to 1M tokens. The main architectural decision is the abandonment of the symmetric decoder-only scheme in favor of an asymmetric Causal Encoder-Decoder (CED). The model is split into a 20-layer encoder that processes the input prompt and a 20-layer decoder that generates the response. The encoder activates 8B parameters per token, while the decoder activates 16B, meaning prompt processing and response generation use different numbers of active parameters.
CED enables substantial memory savings: the decoder’s global KV cache is not computed separately at each decoder layer but is projected from the final hidden state of the encoder. Therefore, when processing a long prompt, the decoder does not recompute its entire history from scratch but restores a short sliding window for the start of generation. As a result, the global KV cache is approximately 890 bytes per token — 4 times less than in DeepSeek-V4-Flash and 437 times less than in DeepSeek-V1. This is the model’s key advantage for long context: KV-cache memory grows significantly more slowly than in comparable decoder-only architectures.
The second pillar of the architecture is Compressed Sparse Attention 2 (CSA2). Instead of the hybrid scheme of different layer types as in V4, here each attention layer is rigidly assigned to one of three modes: Full (computes and stores the main KV and index keys itself, and selects Top-K sparse indices), Reindex (reuses the KV and index keys of the previous Full layer but makes its own Top-K selection), and Reuse (fully reuses both the KV and the already selected indices without performing indexing). Importantly, only a few layers are “full” and bear the main load of building the cache (three layers in the encoder and the first one in the decoder). This scheme allows most layers to be merely users of the shared cache, which reduces both KV-cache memory and computational costs on long context.
Additionally, the model develops several innovative mechanisms. Engram is 196B parameters of conditional memory in the form of n-gram hash tables in layers 1 and 14; they search for patterns in the input and, through a learned gate, are added to the residual stream, acting as a fast associative cache for frequently occurring sequences. The visual part is represented by a 32-layer DeepSeek ViT encoder trained from scratch together with the language model: its embeddings are fed through a two-layer MLP into the text stack, and image tokens retain identifiers, so the MoE router processes them separately without mixing them with text experts. And of course DSpark, which predicts a block from several tokens at once, thereby substantially accelerating generation. These components complement CED and CSA2 and expand the model’s capabilities in multimodal and agentic tasks.
On benchmarks, V4.1-Flash shows strong results and takes first place in a number of tests. On DeepSWE v1.1, the model scores 77.0 versus 73.0 for GPT-5.6 Sol, 74.0 for Claude Opus 5, 67.5 for Kimi K3, and 66.9 for GLM-5.3. On Automation-Bench v1.0.6, the result is 60.5 versus 45.8 for GPT-5.6 Sol, 50.3 for Claude Opus 5, 46.7 for Kimi K3, and 48.8 for GLM-5.3. Other strong indicators include Terminal-Bench 2.1 at 90.6 and GPQA Diamond at 90.9. At the same time, on the most complex science-oriented agentic tasks, a gap with giant models still remains, but on most everyday scenarios the model is quite comparable to closed frontier models.
DeepSeek-V4.1-Flash is well suited for agents and tool loops where large prompts, documents, and the results of numerous tool calls are processed; for software engineering and development automation, including code generation, review, and debugging, work with repositories, and terminal tasks; for automating office processes with tables, documents, and interfaces; for multimodal visual agents with analysis of screenshots, graphs, charts, and layouts; and for a wide range of assistants that need to be able to handle long dialogue history.
| Model Name | Context | Type | GPU | Status | Link |
|---|
There are no public endpoints for this model yet.
An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.
A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.
We recommend creating private endpoints when you need to:
| Name | GPU | TPS | Max Concurrency | |||
|---|---|---|---|---|---|---|
1,048,576.0 pipeline |
3 | $14.38 | 4.726 | Launch | ||
1,048,576.0 tensor |
4 | $19.25 | 16.184 | Launch | ||
1,048,576.0 tensor |
4 | $19.25 | 16.184 | Launch | ||
1,048,576.0 tensor |
8 | $31.39 | 8.844 | Launch | ||
Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.