Deepseek releases V4.1-Flash-1017 open-weight model with improved performance

On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, an upgraded open-weight AI model featuring 552 billion backbone parameters. Trained using reinforcement learning with self-feedback, the model scored 77.0 percent on DeepSWE v1.1 at under $0.20 per task, outperforming larger rivals including Claude Opus 5. DeepSeek also opened its web Search and Fetch APIs to external developers, pricing each service at $0.75 per 1,000 successful requests.

Release and distribution of DeepSeek-V4.1-Flash-1017

• ▪On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, an updated checkpoint of its V4.1-Flash artificial intelligence model accessible via DeepSeek's website, mobile app, and API
• ▪DeepSeek published the open-weight model files for DeepSeek-V4.1-Flash-1017 on Hugging Face and ModelScope on October 17, 2026

API integration and pricing

• ▪DeepSeek enabled developers to access DeepSeek-V4.1-Flash-1017 under the model name deepseek-flash through OpenAI-format or Anthropic-format API endpoints
• ▪DeepSeek set peak API pricing for DeepSeek-V4.1-Flash-1017 at $0.30 per million input tokens and $1.20 per million output tokens, with off-peak rates discounted by 50 percent

Model tuning and performance improvements

• ▪DeepSeek-V4.1-Flash-1017 completes tasks faster and produces shorter outputs than the earlier September 2026 release of DeepSeek-V4.1-Flash
• ▪DeepSeek tuned DeepSeek-V4.1-Flash-1017 for development environments including DeepSeek Harness version v0.2.1-rc.1, Tencent CodeBuddy, and WorkBuddy

Technical architecture and specifications

• ▪DeepSeek-V4.1-Flash features a 1-million token context window, supports up to 384,000 output tokens, and uses 40 Transformer layers divided into a 20-layer causal encoder and 20-layer decoder
• ▪DeepSeek-V4.1-Flash incorporates Compressed Sparse Attention 2 and an FP4 global key-value cache that reduces memory footprint to 890 bytes per token
• ▪DeepSeek-V4.1-Flash architecture utilizes a 552-billion parameter backbone and 196-billion Engram parameters, activating 8 billion parameters per token during prefill and 16 billion during decoding

Reinforcement learning training method

• ▪DeepSeek's reinforcement learning with self-feedback method uses internal confidence scores and self-generated critiques as reward signals, eliminating reliance on human evaluators or external reward models
• ▪DeepSeek trained DeepSeek-V4.1-Flash-1017 using reinforcement learning with self-feedback to improve performance on end-to-end coding, game development, and 3D modeling tasks

Benchmark performance results

• ▪Claude Opus 5 outperformed DeepSeek-V4.1-Flash-1017 on Terminal-Bench 4.0, ProgramBench, NL2Repo-Bench, and Humanity's Last Exam benchmark tests
• ▪DeepSeek-V4.1-Flash-1017 achieved top scores among tested models on Automation-Bench v1.0.6 at 60.5, Agents' Last Exam at 35.4, and Apex Bench v1.1 at 72.1
• ▪DeepSeek-V4.1-Flash-1017 achieved a score of 77.0 percent on the DeepSWE v1.1 benchmark at an average cost under $0.20 per task, surpassing Claude Opus 5's 74.0 percent score

Search API pricing and features

• ▪DeepSeek launched a public Search API endpoint that allows developers to request up to 50 ranked web search results per query from its internal search engine
• ▪DeepSeek set Search API pricing at $0.75 per 1,000 successful requests with a rate limit of 5 queries per second per account

Fetch API features and pricing

• ▪DeepSeek launched a public Fetch API endpoint at POST /v1/tools/fetch that processes single web pages into Markdown format
• ▪DeepSeek priced its Fetch API at $0.75 per 1,000 successful requests, offering free access for server-side calls via select integration tools through late October 2026