DeepSeek V4.1-Flash-1017 Launches With 552B Parameters, 1M Context and $0.003 Input, $0.60 Output Pricing

DeepSeek has introduced its V4.1-Flash-1017 artificial intelligence model, highlighting improvements in benchmark performance, task completion speed and computational efficiency. The model is available through DeepSeek’s website, application and API, according to the company’s announcement.

The latest release focuses on improving AI performance while reducing inference costs and memory requirements. DeepSeek said the model delivers stronger benchmark results, faster task completion and shorter outputs compared with the previous version. It is designed for applications such as AI-powered coding, software development, automated workflows and AI agents.

DeepSeek V4.1-Flash-1017 Benchmark Performance

DeepSeek published benchmark results comparing V4.1-Flash-1017 with models including DeepSeek V4.1-Flash, Kimi K3, GLM 5.3, GPT 5.6-Sol and Claude Opus 5. The company reported a score of 77.0 on DeepSWE v1.1, compared with 74.2 for the previous V4.1-Flash model. On Terminal-Bench 4.0, the new model scored 42.3, compared with 31.2 for the previous version.

The model also recorded 29.1 on ProgramBench, 68.2 on NL2Repo-Bench and 60.5 on Automation-Bench v1.0.6. Its scores on Agents’ Last Exam and Apex Bench v1.1 stood at 35.4 and 72.1, respectively. DeepSeek reported scores of 92.5 on BabyVision and 51.0 on ZeroBench-main.

On the HLE benchmark without tools, DeepSeek V4.1-Flash-1017 scored 42.6 out of 44.0 under the evaluation conditions reported by the company. The published results show that the model improved over its predecessor on several benchmarks, although performance varies across tasks and evaluation methods.

The figures are based on DeepSeek’s published comparisons and should not be interpreted as independent verification of the model’s overall capabilities.

552-Billion-Parameter Mixture-of-Experts Architecture

DeepSeek V4.1-Flash-1017 uses a mixture-of-experts (MoE) architecture with 552 billion parameters. According to the company, its causal encoder-decoder design activates approximately 8 billion parameters for input processing and 16 billion parameters for output generation.

The architecture is designed to reduce the amount of computation required for individual operations while retaining the capacity of a large model. By activating a smaller subset of parameters for specific tasks, the model aims to improve inference efficiency and reduce operating costs.

DeepSeek also attributed the model’s performance to updated pretraining methods and reinforcement learning-based post-training. The company said these changes helped the Flash model achieve benchmark results ahead of several competing models on selected evaluations.

DeepSeek Reduces KV Cache Requirements

The company has also highlighted a reduction in key-value (KV) cache requirements, which can affect memory consumption and the cost of processing long conversations or repeated prompts.

According to DeepSeek’s published figures, the global KV cache requirement has fallen from 389,120 bytes per token in DeepSeek-V1 to 48,068 bytes in DeepSeek-V3.2. The requirement declined further to 3,514 bytes in DeepSeek V4.1-Flash and approximately 890 bytes in V4.1-Flash-1017.

This represents a substantial reduction compared with earlier model generations. DeepSeek said the V4.1-Flash architecture requires only one-quarter of the high-bandwidth memory (HBM) and one-eighth of the solid-state drive (SSD) storage required by the previous generation, according to its comparison.

Smaller KV caches can reduce memory consumption and improve the economics of AI inference, particularly for applications that process long context windows or make repeated requests. The actual savings depend on deployment configurations, workload characteristics and caching behaviour.

DeepSeek V4.1-Flash API Pricing

DeepSeek has introduced separate API rates for cached input tokens, uncached input tokens and output tokens, with prices varying according to peak and off-peak periods.

During off-peak hours, cached input tokens cost $0.003 per million tokens, while uncached input tokens cost $0.15 per million tokens. Output tokens are priced at $0.60 per million tokens during the same period.

During peak hours, the price increases to $0.006 per million cached input tokens, $0.30 per million uncached input tokens and $1.20 per million output tokens. DeepSeek said off-peak rates are 50% lower than peak rates.

Peak hours are defined as 01:00 to 04:00 UTC and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. All other periods, including weekends and Chinese public holidays, are classified as off-peak.

The pricing structure allows developers to estimate costs according to input volume, cache-hit rates and generated output. Applications that can schedule non-urgent workloads during off-peak hours may benefit from the lower rates.

API Features and Model Availability

DeepSeek V4.1-Flash-1017 is available through the company’s API using the model identifier deepseek-flash. The published model documentation lists a context length of one million tokens and a maximum output of 384,000 tokens.

The model supports structured JSON output, tool calls, the Responses API and Anthropic-compatible API access. It also supports vision capabilities, allowing developers to build applications that process both text and visual information.

DeepSeek stated that the older model identifiers deepseek-v4-flash and deepseek-v4-flash-vision-exp would continue to be accepted for compatibility. Requests using these legacy names would be routed to the newer V4.1-Flash model and billed at the applicable Flash rates.

The company also announced compatibility with development platforms and tools, including WorkBuddy, CodeBuddy and OpenCode. Developers using DeepSeek Harness were advised to upgrade to version 0.2.1-rc.1 or later for the recommended experience.

DeepSeek Expands Search and Fetch API Access

Alongside the model release, DeepSeek announced broader access to its Search and Fetch APIs, which use the company’s in-house search engine to retrieve web information.

The published price for each service is $0.75 per 1,000 successful requests. Search and Fetch are billed separately, meaning developers using both services may incur charges for each endpoint.

DeepSeek also stated that Search and Fetch calls made through supported server-side tools in DeepSeek Harness, Claude Code and Codex would remain free for DeepSeek API users until October 31, 2026, at 16:00 UTC. The company said billing for these calls would begin after that date.

These services can be used to incorporate web search and content retrieval into AI applications, including research assistants, coding agents and automated information-processing systems.

Open-Source Support and Large-Scale Deployment

DeepSeek said it plans to work with the open-source community to expand inference support and deployment options for V4.1-Flash. The company has also made the model’s technical report and model files available through the resources linked in its announcement.

The company invited organisations planning large-scale deployments involving 2,000 GPUs or more, alongside storage infrastructure, to discuss their requirements with DeepSeek.

The release combines benchmark improvements with reduced cache requirements and differentiated API pricing. These changes could make the model relevant to developers building coding agents, software engineering tools and other applications that process large volumes of tokens.

However, the overall cost and performance of V4.1-Flash-1017 will depend on the specific workload, inference configuration, cache-hit rate and output volume. Developers evaluating the model against alternatives will need to compare both task accuracy and total operating costs under comparable conditions.