- Chat
- Vision
DeepSeek V4.1 Flash
Multimodal mixture-of-experts model: 552B backbone parameters, with 8B active during prefill and 16B during decoding.
DeepSeek model family
About this model
DeepSeek V4.1 Flash is a multimodal mixture-of-experts model for image and text input with text output. Its causal encoder-decoder has 40 layers: a 20-layer encoder and a 20-layer decoder. The published model has 552B backbone parameters, activates 8B during prefill and 16B during decoding, and includes a separate 196B-parameter Engram conditional memory. It uses Compressed Sparse Attention 2, Single-Pass mHC and DSpark speculative decoding. The upstream configuration supports 1,048,576 context tokens and reasoning effort from 1 to 100. Coming soon to Arnict; serving limits and pricing will be published at launch.
- Parameters
- 552B552B backbone parameters; separate 196B Engram conditional memory.
- Native context
- 1,048,5761,048,576-token upstream configuration; Arnict serving limit pending launch.
- Output guidance
- —Published guidance
Capabilities
- Chat completionsUse this model with the chat-completions API when it is marked Available.
- Native input modalitiesText, Image input are listed in the published catalog metadata.
- Long-context input1,048,576 tokens · 1,048,576-token upstream configuration; Arnict serving limit pending launch.
Published limits
Published model-card facts, kept separate from live availability.
| Metric | Published value | Note |
|---|---|---|
| Arnict API output limit | Not published | Maximum output accepted by the Arnict API for this model |
| Parameters | 552B backbone parameters; separate 196B Engram conditional memory. 8B prefill / 16B decode active. | Upstream model card |
| Context window | 1,048,576 tokens | 1,048,576-token upstream configuration; Arnict serving limit pending launch. |
| Output guidance | Not published by upstream | Arnict output limit will be published at launch. |
| Weights / activations | FP4 expert weights / FP8 activations (upstream) | Published format |
- 40 layers: 20-layer causal encoder + 20-layer decoder
- 384 routed experts; 6 selected per token; 1 shared expert
- Compressed Sparse Attention 2 (CSA2) and Single-Pass mHC
- Engram conditional memory and DSpark speculative decoding
- Native image and text input; text output
- Upstream reasoning effort: integer 1-100
Use deepseek/deepseek-v4.1-flash with an OpenAI-compatible client after this model is enabled.
Related models
DeepSeek
DeepSeek V4 Pro 0813
DeepSeek V4 Pro mixture-of-experts model with DSpark speculative decoding and a 1,048,576-token context.
GLM
GLM 5.3 Flash
Long-context chat, coding and tool workflows with GLM 5.3 Flash.
Qwen
Qwen 3.6 35B A3B
Vision-language mixture-of-experts model: 35B total, 3B active parameters.