DeepSeek launched an improved AI model with better coding scores and opened its search tools to all developers for immediate use.
Key points
- The new DeepSeek-V4.1-Flash-1017 checkpoint scores 77.0 on DeepSWE v1.1, outperforming Claude Opus 5 in that specific benchmark.
- DeepSeek opened its Search and Fetch APIs to every developer, allowing agent builders to access web data through a single API key.
- The model uses a Mixture-of-Experts architecture with 552B backbone parameters, activating only 8B to 16B parameters per token during processing.
DeepSeek has released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model, alongside open access to its Search and Fetch APIs. This update aims to provide faster task completion and stronger performance scores compared to the previous September release, while reducing costs for developers building AI agents.
What happened
According to MarkTechPost, DeepSeek made DeepSeek-V4.1-Flash-1017 live on its website, app and API. The company describes this release as a Pareto improvement, meaning it offers better performance and lower cost than the September V4.1-Flash version. The model beats its predecessor on all eleven benchmarks reported by DeepSeek, with gains ranging from 2.0 to 11.1 points. It achieves a score of 77.0 on DeepSWE v1.1, which is higher than the 74.0 score for Claude Opus 5 in DeepSeek’s comparison table. However, it scores 42.6 on HLE without tools, compared to 56.3 for Claude Opus 5.
View the original post on X
The model features a multimodal Mixture-of-Experts architecture with 40 Transformer layers. It has 552B backbone parameters and 196B Engram parameters, but activates only 8B parameters per token during prefill and 16B during decode. It supports a context window of 1 million tokens with up to 384K output tokens. DeepSeek also opened its Search and Fetch APIs to all developers. This allows agent builders to use a cheaper coding model combined with a first-party web layer behind one API key. The September V4.1-Flash weights are MIT-licensed on Hugging Face, but hardware requirements and weights for the 1017 version are not disclosed.
Why it matters
The fix
- Check if your current AI agent workflows can integrate with the newly opened DeepSeek Search and Fetch APIs.
- Review your API usage costs, as DeepSeek claims this model offers a cost improvement over previous versions.
- Verify that your development environment supports the recommended DeepSeek Harness v0.2.1-rc.1 or later for optimal performance.
- Test the model’s output length and task completion speed against your previous benchmarks to confirm the claimed improvements.
Why it matters
Understanding this release requires knowing how Mixture-of-Experts models work. These systems use many small, specialized neural networks instead of one large model. Only the relevant experts activate for each input, which reduces compute costs and speeds up processing. This architecture allows developers to run powerful models on less hardware. Opening search APIs means developers can build agents that fetch real-time data without building their own web scraping infrastructure. Teams usually evaluate such updates by testing specific benchmarks relevant to their workflow, such as coding accuracy or reasoning speed, rather than relying solely on aggregate scores.
What to do and how to stay safe: DeepSeek
- Update your API keys and configuration files to point to the new deepseek-flash model endpoint.
- Test the model’s performance on your specific coding or reasoning tasks to verify the claimed improvements.
- Monitor your API usage and costs to ensure the Pareto improvement claims hold true for your workload.
- Review the MIT license terms for the September V4.1-Flash weights if you plan to host models locally.
Step-by-step guide: What is password policies and how they protect systems
General security guidance from the Crypt Weekly newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does DeepSeek-V4.1-Flash-1017 compare to Claude Opus 5?
DeepSeek reports a score of 77.0 on DeepSWE v1.1, which is higher than Claude Opus 5’s 74.0, but lower on HLE without tools at 42.6 versus 56.3.
Is the model available for local deployment?
The September V4.1-Flash weights are MIT-licensed on Hugging Face, but hardware requirements and weights for the 1017 checkpoint are not disclosed.
What new APIs are available to developers?
DeepSeek opened its Search and Fetch APIs to every developer, allowing integration with the model through a single API key.
