Chinese AI startup DeepSeek has released DeepSeek-V4.1-Flash, which it describes as the smallest model in a new architecture family. The launch comes as the company is reportedly preparing for an initial public offering on Shanghai’s technology-focused STAR Market.

The multimodal model has a 552-billion-parameter mixture-of-experts backbone and supports context windows of up to one million tokens. Its main advance is efficiency rather than sheer scale. DeepSeek’s causal encoder-decoder architecture divides 40 transformer layers into separate 20-layer encoding and decoding stages. This means that eight billion parameters are activated for each input token and 16 billion for each generated token.

The architecture targets input-heavy agentic tasks, in which processing large documents, codebases or conversation histories can account for a significant share of computing costs. DeepSeek said its SWA Bounded Replay method eliminates the need to persist selected attention states on solid-state storage, reducing the persistent KV-cache footprint to approximately one-eighth of that used by DeepSeek-V4-Flash.

A related Compressed Sparse Attention 2 system assigns static attention modes across layers and reuses selected attention indices. Combined with FP4-formatted key-value caching, the system reduces the global KV-cache requirement to 890 bytes per token, or about one-quarter of the previous Flash model’s requirement. These figures describe memory used to retain context during inference, separate from the model’s total parameter count. V4.1-Flash also incorporates conditional memory and speculative decoding, and uses one shared expert and 384 routed experts per MoE layer.

DeepSeek’s official website is and the model is available at

Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. Introducing the smallest model in our new architecture family, with native visual understanding. Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6 pic.twitter.com/wxJGiyX56o — DeepSeek (@deepseek_ai) September 10, 2026

The announcement is also available on X: and https://x.com/deepseek_ai/status/2097930608790167907?ref_src=twsrc%5Etfw.

Competitive Results Across Reasoning, Coding and Agents

DeepSeek trained V4.1-Flash from scratch on 45 trillion multimodal tokens, with images processed natively alongside text from the beginning of pretraining. Sparse-attention training began at 64,000 tokens, before the context length was extended to one million tokens at the 34-trillion-token stage. Post-training combined supervised fine-tuning, reinforcement learning and on-policy distillation. Reasoning effort can be configured on a scale from one to 100.

The reported benchmarks show strong performance relative to much larger models. The base version scored 74.1 on MMLU-Pro, 79.4 on HumanEval and 93.0 on GSM8K, compared with scores of 73.5, 76.8 and 92.6 for DeepSeek-V4-Pro-Base. At maximum reasoning effort, V4.1-Flash scored 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, placing it slightly ahead of leading Opus and GPT models on those agent benchmarks. It also recorded a 3,471 Codeforces rating and matched the best listed result on MathArena Apex.

The results were not uniformly dominant. V4.1-Flash remained behind the largest comparison models on GPQA Diamond, Humanity’s Last Exam and SimpleQA. Its scores on the newer Terminal-Bench 3.0 and 4.0 evaluations also trailed Opus-5.0, although they represented substantial gains over earlier DeepSeek models.

Multimodal results included scores of 56.5 on MMMU-Pro, 77.9 on CVBench and 95.6 on DocVQA. With agent tools enabled, the model scored 78.9 on Chartography and 89.6 on BabyVision.

The original report was published by Metaverse Post: https://mpost.io/new-deepseek-v4-1-flash-challenges-larger-ai-systems-on-coding-and-agent-tasks/