New DeepSeek V4.1-Flash Challenges Larger AI Systems On Coding And Agent Tasks

Chinese AI startup DeepSeek released DeepSeek-V4.1-Flash, describing it 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 contexts of up to one million tokens. Its principal advance is efficiency rather than sheer size. DeepSeek’s causal encoder-decoder architecture divides 40 transformer layers into separate 20-layer encoding and decoding stages. As a result, only eight billion parameters are activated for each input token and 16 billion for each generated token.

The architecture is designed for input-heavy agentic tasks, where processing large documents, codebases or conversation histories can dominate computing costs. DeepSeek said its SWA Bounded Replay method removes the need to persist selected attention states on solid-state storage, reducing the persistent KV-cache footprint to approximately one-eighth that of DeepSeek-V4-Flash.

A related Compressed Sparse Attention 2 system assigns static attention modes across layers and reuses selected attention indices. In combination with FP4-formatted key-value caching, it reduces the global KV-cache requirement to 890 bytes per token, about one-quarter of the previous Flash model. The model also incorporates conditional memory and speculative decoding while using one shared and 384 routed experts per MoE layer.

Competitive Results Across Reasoning, Coding and Agents

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

Benchmarks indicate 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 73.5, 76.8 and 92.6 for DeepSeek-V4-Pro-Base. At maximum reasoning effort, V4.1-Flash achieved 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 posted a 3,471 Codeforces rating and matched the best listed result on MathArena Apex.

The results are 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 also trailed Opus-5.0, despite representing substantial gains over earlier DeepSeek models.

Multimodal performance included 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 post New DeepSeek V4.1-Flash Challenges Larger AI Systems On Coding And Agent Tasks appeared first on Metaverse Post.