DeepSeek V4.1-Flash

Generic Info

• Publisher: DeepSeek AI
• Release Date: September 2026
• Parameters: 552B Backbone (8B active during prefill, 16B during decode)
• Architecture: Causal Encoder-Decoder (CED) + Compressed Sparse Attention 2 (CSA2)
• Context Window: 1,000,000 tokens (1M tokens)
• License: MIT License
• Key Capabilities: 1/8 Persistent KV Footprint, Native Multimodal (Vision+Text), Ultra-Low Latency Agentic Prefill

DeepSeek-V4.1-Flash introduces a revolutionary Causal Encoder-Decoder (CED) architecture organized as a 20-layer causal encoder followed by a 20-layer decoder. By projecting the decoder's global KV cache from the final encoder hidden states, the model activates only 8B parameters per token during prefill and 16B during decode. Combined with SWA Bounded Replay, it slashes the persistent KV cache footprint to roughly 1/8th of conventional architectures, unlocking blazingly fast inference on massive document inputs.

Hello World Guide

Deploy DeepSeek-V4.1-Flash locally or connect via the OpenAI-compatible API endpoint.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "deepseek-ai/DeepSeek-V4.1-Flash"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "Explain how the Causal Encoder-Decoder (CED) structure optimizes KV memory during 1M token prefill."
messages = [
    {"role": "system", "content": "You are DeepSeek-V4.1, an ultra-fast agentic reasoning model."},
    {"role": "user", "content": prompt}
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.6
)

response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

Industry Usage