DeepSeek-V4.1-Flash 在部分智能体与编程测试中超越 GPT-5.6 Sol

DeepSeek has unveiled its latest artificial intelligence model, DeepSeek-V4.1-Flash. Although it has fewer parameters than DeepSeek's top-tier V4-Pro model, the new model showed advantages in performance, cost and speed that could make it a practical substitute for the flagship. It also outperformed V4-Pro on some agentic and coding benchmarks.

DeepSeek-V4.1-Flash sharply reduces not only the number of active parameters but also its key-value (KV) cache, while posting agentic and coding performance ahead of V4-Pro and, on some benchmarks, several frontier models. DeepSeek said Monday that it would begin consolidating its existing V4-Pro application programming interface (API) service into V4.1-Flash.

DeepSeek-V4.1-Flash uses a Mixture of Experts (MoE) architecture, which activates only selected expert networks depending on the input to improve computational efficiency. The model has 552 billion parameters in its backbone, with 8 billion active parameters during input processing and 16 billion during output generation.

The previous V4-Flash had 284 billion parameters and 13 billion active parameters. The top-end V4-Pro had 1.6 trillion parameters and 49 billion active parameters. Despite that difference, DeepSeek says V4.1-Flash delivers better performance than V4-Pro in agentic and coding tasks. The distinction between active parameters for input and output in DeepSeek-V4.1-Flash is particularly notable.

DeepSeek-V4.1-Flash is designed with a focus on agents and coding. According to DeepSeek, at the maximum reasoning intensity, or reasoning_effort=100, V4.1-Flash outperformed OpenAI's GPT-5.6 Sol and Anthropic's Claude Opus 5 on Terminal-Bench 2.1. DeepSeek also claims an advantage over the two models on DeepSWE v1.1 and AutomationBench.

DeepSeek-V4.1-Flash performance comparison with major frontier models on agentic benchmarks. (Source: Hugging Face model card)

Outside agentic and coding tasks, however, the model posted relatively lower scores on difficult general reasoning. Artificial Analysis (AA), an independent AI model evaluation platform, gave the model an Intelligence Index score of 40. That was well above the average for open-weight models but below the leading frontier models.

AA calculates its Intelligence Index by combining multiple evaluations, including GDPval-AA, AutomationBench-AA, SciCode, Humanity's Last Exam (HLE), CritPt and AA-LCR. DeepSeek-V4.1-Flash recorded scores including 1,632 on GDPval-AA v2, 69% on AutomationBench-AA, 52% on SciCode and 39% on HLE, leaving the model in the upper-middle range overall. The results indicate that the model can be highly efficient when deployed for coding-focused applications.

The score of 40 also reflects AA's recent changes to its evaluation methodology, which have lowered the overall scale of the index. AA upgraded its Intelligence Index from v4.2 to v4.3, making the evaluation set more difficult.

As a result, frontier models that previously scored well above 60 have been moved into the 50s under the revised methodology. Anthropic's Claude Fable 5.1, which had scored 66 under the previous v4.2 evaluation, now scores 53 under the revised system, alongside OpenAI's GPT-6 Astra.

DeepSeek's performance comparison between V4.1-Flash and other AI models, highlighting the model's claimed advantage in coding. (Source: Hugging Face model card)

AA's upgrade was driven in part by the replacement of Terminal-Bench 2.1 with Terminal-Bench 4.0. DeepSeek-V4.1-Flash outperformed GPT-5.6 Sol and Claude Opus 5 on Terminal-Bench 2.1, but fell behind them on Terminal-Bench 4.0.

Terminal-Bench 4.0 consists of 66 complex terminal tasks spanning software, machine learning, science, operations, security, hardware and media. It revised computing and time allowances and improved task descriptions, environments and verification logic to more finely distinguish performance among frontier models.

In terms of parameter size, DeepSeek-V4.1-Flash's Intelligence Index score is slightly below the 42 points recorded by GLM-5.3-Flash but is well above the median score of 18 for other open-weight models of a similar size. Under AA's stricter Intelligence Index, GLM-5.3-Flash ranks first among open-weight models, while Alibaba's Qwen 3.8 Max and DeepSeek-V4.1-Flash share second place at 40. DeepSeek V4 Pro 0813 scored 36.

DeepSeek-V4.1-Flash is also fast, as DeepSeek claims. Its output speed is about 214.4 tokens per second, while time to first token (TTFT) is 1.13 seconds. The median for comparable models is 65.4 tokens per second and 2.34 seconds for TTFT, making V4.1-Flash more than three times as fast by the output-speed measure. AA also rated it as “notably fast.”

Artificial Analysis' overall evaluation of DeepSeek-V4.1-Flash, including intelligence, speed, cost and verbosity metrics. (Source: Artificial Analysis)

The model has reduced its KV cache to 890 bytes per token. AA also viewed this as a significant strength. The reduction becomes more consequential when combined with DeepSeek-V4.1-Flash's pricing structure.

According to AA, DeepSeek-V4.1-Flash is priced competitively for input and output tokens, with an overall cost below the median of comparable models. Cache-hit pricing receives further reductions, while off-peak periods receive an additional 50% discount.

AA, however, said the model is “very verbose.” Its Intelligence Index evaluation used 250 million output tokens, substantially more than the 130 million median for similar open-weight models. The total cost of the evaluation was also recorded. Although the model is fast, the length of its responses can affect actual cost and completion time.

Still, AA said the model reaches the Pareto frontier on a cost-per-task basis. In other words, it is among the least expensive models at a comparable level of intelligence.

Artificial Analysis' AI model evaluation as of Sept. 14, with DeepSeek-V4.1-Flash indicated by the red arrow; lower task cost is better. (Source: Artificial Analysis)

The reason DeepSeek-V4.1-Flash separates active parameters for input and output, unlike a conventional MoE model, lies in its Causal Encoder-Decoder (CED) architecture.

The term CED, or the concept behind it, was not created by DeepSeek. The term has been used since the late 2010s in fields including speech, time-series modeling and generative models. A paper on Alibaba's Qwen-Audio-VAE, published in July, also discusses CED. However, the CED used by DeepSeek in V4.1-Flash appears to be a different technology despite sharing the same name.

DeepSeek's CED makes the encoder causal rather than bidirectional and eliminates the need for cross-attention by projecting the encoder's final representation into the decoder's KV cache in a single operation. The structure appears to have been designed and named by DeepSeek for V4.1-Flash, and DeepSeek identifies it as a new architecture for its model in the official Hugging Face model card and technical documentation.

DeepSeek's CED in V4.1-Flash does three things: it changes the originally bidirectional encoder into a causal, sequential structure so it follows the same basic pretraining pattern used by today's autoregressive models; it eliminates the need for the decoder to query the encoder repeatedly through cross-attention by transferring the encoder's final output into the decoder cache in one operation; and it repurposes the architecture as a cost- and memory-saving mechanism inside a conversational chatbot.

Artificial Analysis' upgrade from Terminal-Bench 2.1 to Terminal-Bench 4.0; DeepSeek-V4.1-Flash led frontier models on Terminal-Bench 2.1 but ranked lower on Terminal-Bench 4.0. (Source: Artificial Analysis)

According to DeepSeek's technical report, the 40 Transformer layers in V4.1-Flash are divided into 20 causal encoder layers and 20 decoder layers. Instead of generating global KV separately at each decoder layer, the model projects the decoder's global KV from the final hidden state of the encoder.

DeepSeek's CED architecture does not correspond directly to either the original Transformer encoder-decoder architecture or the decoder-only architecture used by most modern large language models (LLMs). According to DeepSeek's technical documentation, it differs in three key ways.

First, the encoder itself is causal. In the original encoder-decoder architecture, the encoder reads a sentence bidirectionally, looking both forward and backward. In DeepSeek's CED, the encoder also reads sequentially, from left to right. This is why it is called a causal encoder. Making the encoder causal in the same way as the decoder allows all 40 layers to operate under a consistent causal scheme, simplifying training.

Second, DeepSeek eliminates cross-attention, through which the decoder would otherwise query information from the encoder in real time at every layer and for every generated token. Cross-attention is an attention mechanism in which one part of a model references information held by another part rather than relying only on its own representations.

How DeepSeek CED Differs From Encoder-Decoder and Decoder-Only Models
Early Transformer-based models used an encoder-decoder architecture, while most generative large language models (LLMs) today use a decoder-only architecture. In an encoder-decoder model, dozens of layers are stacked sequentially. When a user submits a question or instruction, the input passes through the layers from the first to the last, becoming progressively more refined and processed. The process of understanding the user's prompt and the process of generating a response are handled through the same architecture, requiring the model to pass through all of those layers during processing. If a user attaches a very long document, or an AI agent performs a continuous sequence of tasks over an extended period as the conversation history keeps growing, the AI must repeatedly process that information through dozens of layers. Each layer leaves key-value (KV) cache states to retain important information. In general, the more layers a model has, the more KV states need to be stored. That is why KV cache is considered a major bottleneck and a significant source of rising costs in AI service operations. This traces back to the Transformer introduced by Google researchers in 2017, when machine translation was presented as a key application. In an English-to-Korean translation example, the architecture can broadly be divided into an encoder, which reads the entire English source sentence, and a decoder, which uses that understanding to generate the Korean translation. The key point is that the encoder reads the sentence bidirectionally. For example, when processing the fifth word of a sentence, the encoder can refer not only to the first four words but also to all words after the sixth. This is similar to a person reading an entire sentence before determining its context. Once the encoder has fully processed the source sentence in both directions, the decoder uses that representation to generate the translation one word at a time. The decoder operates only from left to right, or causally. To generate the next word, it can refer to words it has already produced, but cannot see words that have not yet been generated. When generating each word, however, the decoder does more than look at the words it has previously produced. It also references the representations generated by the encoder from the input tokens. This process is known as cross-attention. Crucially, cross-attention occurs at every decoder layer and each time a word is generated. In other words, whether at the fifth decoder layer or the 20th, the decoder repeatedly references information produced by the encoder in real time. This architecture is useful for tasks such as translation and summarization, where the input and output may be in different languages or have substantially different formats. Today's widely used models, including GPT and Claude, generally use a decoder-only architecture rather than the traditional encoder-decoder structure. The encoder and decoder are not separated. Instead, the user's question and the AI's response are combined into a single long sequence of text, which is processed by one set of layers entirely in a causal, left-to-right manner. The process of understanding the question and the process of generating the answer therefore share the same layers and parameters. The architecture is relatively simple and efficient to train, helping decoder-only models become an industry standard following the GPT series. Its drawback is that a long input must pass through all layers when it is initially processed to generate the KV cache, while the memory burden of storing and managing the cache increases as conversations grow longer. DeepSeek's Causal Encoder-Decoder (CED) architecture attempts to reduce this repeated processing. The basic idea is relatively straightforward: instead of treating all 40 workstations as a single processing line, the model divides them into two groups of 20 with distinct roles. The first 20 workstations, or encoder layers, process the input and produce the final hidden states. The user's question or document is read once and transformed into a representation that can be used by subsequent computations. The second 20 workstations, or decoder layers, generate the response based on that information and construct the KV states needed during generation.

Third, the roles of the encoder and decoder are changed. The model separates input understanding from response generation while continuing to operate as a single conversational chatbot from the user's perspective. Internally, the input-processing portion activates 8 billion parameters, while the output-generation portion activates 16 billion.

Early Transformer-based models used an encoder-decoder architecture, while most current generative LLMs use a decoder-only architecture.

GDPval-AA evaluation of DeepSeek-V4.1-Flash, measuring performance on complex professional tasks across industries and occupations. (Source: Artificial Analysis)

SWA and DeepSeek’s SWA Bounded Replay
AI models broadly rely on two types of information when generating a response: global memory and local memory. Global memory maintains information across a long context. In DeepSeek's Causal Encoder-Decoder (CED) architecture, the encoder processes the entire input and uses its final hidden states to generate the decoder's global key-value (KV) cache. This global KV cache maintains an overall representation of the long context. But when a large language model (LLM) generates a response, it needs more than an understanding of the entire context. It also needs fine-grained information within a very narrow context, such as which word it has just used or how the preceding sentence ended. Information from the full long context alone may not adequately preserve such details, making a separate form of local memory necessary. Sliding Window Attention (SWA) is a mechanism that allows tokens within only a recent portion of the sequence to attend to one another. In DeepSeek's CED architecture, however, the decoder does not retain its SWA local state in full. DeepSeek therefore uses a separate mechanism called SWA Bounded Replay to compensate for the missing local state. In DeepSeek's CED architecture, the decoder does not repeatedly process the entire prompt as conventional decoder layers would. The first 20 encoder layers process the full prompt and produce the final hidden states, which are then used to construct the global KV cache for the following 20 decoder layers. The decoder can therefore access information from the entire input processed by the encoder through the global KV cache. That does not mean the model lacks local context information. Response generation still requires fine-grained information from a narrow context, such as the word just generated or the immediately preceding sentence. DeepSeek-V4.1-Flash addresses this by maintaining SWA state only for the most recent 128 tokens. Rather than continuously storing local KV on a solid-state drive (SSD), the model can reprocess the most recent 128 tokens when needed to reconstruct the local state. This is the role of SWA Bounded Replay. With SWA Bounded Replay, DeepSeek-V4.1-Flash performs a short replay process in which only the final 128 tokens of the prompt pass through the 20 decoder layers. This allows the decoder to recover local memory, or SWA state, for the immediately preceding context. The process is called “bounded” because the replay is limited to the most recent 128 tokens rather than the full prompt, and “replay” because the decoder recalculates the information to recover the recent context. The approach is a key part of DeepSeek's effort to improve cost efficiency. Under a conventional decoder architecture, preserving local memory would require an SWA cache containing recent-context information to be retained for the full length of the prompt. As conversations become longer and prompts grow, the amount of storage required also increases, eventually creating a need to move the cache to disk or an SSD. SWA Bounded Replay significantly reduces the need for such local-memory storage. When the local state is needed, just before response generation, the system can immediately recalculate the most recent 128 tokens, use the resulting state and then discard it. Instead of storing the SWA KV state on an SSD, the system reconstructs it by replaying the most recent 128 tokens when required. DeepSeek says this approach eliminates persistent SSD storage for SWA KV and, as a result, reduces the overall persistent KV cache footprint to about one-eighth that of the previous V4-Flash. While CED's reuse of global memory reduces computation and the amount of global cache that needs to be maintained, SWA Bounded Replay serves as a complementary mechanism that fills the structural gap in local memory at a small, fixed computational cost while also reducing storage requirements.

Another technology used to reduce the KV cache is △Compressed Sparse Attention 2 (CSA2), △along with FP4 quantization. CSA2 allows multiple layers to share and reuse KV and attention indices, reducing the duplication that would otherwise occur when each layer maintains its own cache. FP4 quantization stores the global KV cache using four-bit floating-point values, reducing storage requirements compared with FP8.

DeepSeek introduced the first version of Compressed Sparse Attention (CSA) in V4. While DeepSeek Sparse Attention (DSA) selects important KV entries and performs attention only on a subset of them in long contexts, CSA first compresses the KV representations of multiple tokens into shorter representations and then sparsely selects the information it needs. In other words, CSA combines KV compression with sparse selection to reduce both KV cache size and attention computation for long contexts.

V4.1-Flash advances the approach to CSA2. Each attention layer is assigned one of three modes — Full, Reindex or Reuse — allowing multiple attention layers to share and reuse KV and the locations of important information rather than recalculating the same information at every layer. Some layers perform new calculations while others reuse existing results, reducing the memory and computation required for long contexts. Combined with FP4 KV caching, the global KV cache is reduced to 890 bytes per token, about one-quarter the size of the V4-Flash cache.

Reducing the number of active parameters while maintaining performance lowers the GPU memory and computational resources required to run an AI model. That can reduce server operating and power costs and allow the same GPU resources to handle more simultaneous requests. Lower computational requirements can also reduce latency and make deployment on lower-end GPUs, on-premises servers rather than cloud infrastructure, and edge AI systems easier.

AutomationBench-AA evaluation of DeepSeek-V4.1-Flash, which measures an AI agent's ability to integrate with business software and complete automated tasks. (Source: Artificial Analysis)

If total parameters can be reduced while maintaining the same performance, the model's weight memory can also be reduced, which is the ideal outcome. If that is not possible, reducing the active parameters in an MoE model is the next-best option because fewer active parameters also mean fewer computations.

DeepSeek-V4.1-Flash has a total of 552 billion parameters, which is substantial, but activates only 8 billion parameters for input and 16 billion for output. CSA2 also sharply reduces the KV cache, allowing GPU memory to be used more efficiently. The fact that DeepSeek developed and actively deployed technologies such as CSA2 and CED in the model is particularly notable.

DeepSeek is not alone. Chinese AI companies including Moonshot AI, developer of the Kimi family, and Alibaba, developer of Qwen, are expanding model parameter counts or advancing their models through proprietary technologies. Questions remain, however, over allegations that some Chinese AI companies illegally distilled capabilities from U.S. frontier models.

Anthropic recently said that seven Chinese AI companies had attempted to extract capabilities from its Claude models without authorization, and disclosed cases involving Alibaba, Moonshot AI and DeepSeek.

According to Anthropic's findings, Alibaba linked more than 151 million interactions with Claude to its own models. Anthropic said Alibaba created large numbers of accounts to send queries to Claude and collect its responses, then used the data to train or improve its own AI models. In effect, Claude's responses were used as teacher-model data.

In Moonshot AI's case, more than 23 million interactions were identified. Particularly problematic, according to Anthropic, was that Moonshot forwarded questions entered by actual Kimi users to Claude and then displayed Claude's responses to those users as if they had been generated by Kimi. Anthropic also said Moonshot subsequently used the collected user questions and Claude responses in developing its own models.

For DeepSeek, Anthropic said it identified about 12.1 million interactions. Anthropic alleged that DeepSeek also sent large numbers of queries to Claude to collect responses and used them for model distillation.