DeepSeek releases V4.1-Flash-1017 model with improved coding benchmarks
The new open-weight model achieves higher scores on coding tasks and includes an updated Search API for developers.
Key points
- DeepSeek-V4.1-Flash-1017 scores 77.0 on DeepSWE v1.1, surpassing previous versions.
- The model uses reinforcement learning with self-feedback to improve agentic tasks.
- DeepSeek opened its Search and Fetch APIs to all developers alongside the release.
DeepSeek has released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model, making it available on its website, app, and API. The company claims this version offers stronger benchmark scores, faster task completion, and shorter outputs compared to the September V4.1-Flash release. This follows earlier reporting by Threat Times on October 18, which noted the model's focus on reinforcement learning for agentic tasks.
According to MarkTechPost, the model is described by the company as a Pareto improvement in both performance and cost. It activates eight billion parameters per token during prefill and sixteen billion during decode. The architecture supports a context window of one million tokens. Hardware requirements and the specific weights for the 1017 checkpoint have not been disclosed by the vendor at this time.
View the original post on X
Model Performance and Architecture
DeepSeek states that V4.1-Flash-1017 beats the previous V4.1-Flash release on all eleven benchmarks it reported, with gains ranging from 2.0 to 11.1 points. The highest score achieved was 77.0 on DeepSWE v1.1, which exceeds the 74.0 score for Claude Opus 5 in DeepSeek’s comparison table. However, the model scored 42.6 on HLE without tools, compared to 56.3 for Claude Opus 5.
The underlying model is a multimodal Mixture-of-Experts system with 40 Transformer layers, split into a 20-layer causal encoder and a 20-layer decoder. It utilizes a Causal Encoder-Decoder architecture where decoder global key-value is projected from final encoder hidden states, nearly halving prefill compute. It also employs Compressed Sparse Attention 2 and FP4 main key-value cache to reduce memory usage.
Developer Access and API Updates
Alongside the model release, DeepSeek opened its Search and Fetch APIs to every developer. This move allows agent builders to access a cheaper coding model and a first-party web layer behind a single API key. The model is served through the deepseek-flash name on the DeepSeek API. The September V4.1-Flash weights, comprising 763 billion parameters in FP8 format, are available under an MIT license on Hugging Face.
The new checkpoint is tuned for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. The DeepSeek team recommends using DeepSeek Harness v0.2.1-rc.1 or later. The model uses reinforcement learning with self-feedback for end-to-end project work, with examples ranging from game creation to 3D modeling. Reasoning effort is adjustable via a scalar from 1 to 100, with API tiers mapping to 50, 75, and 100.
What should you do?
Security professionals and developers should review the terms of service and privacy policies for the newly opened Search and Fetch APIs before integrating them into production applications. Teams using the DeepSeek API should monitor the deepseek-flash endpoint for any changes in latency or output behavior as the new checkpoint is rolled out. Organizations relying on the September weights should note that hardware requirements for the new 1017 checkpoint are not yet public, which may impact local deployment plans.
Background: Privacy policies
A privacy policy is a legal document that explains what data a company collects, why it collects it, and how it shares that information. It defines your rights regarding consent and deletion. Understanding it helps you control your digital footprint and avoid unintended data exposure to third parties.
Read the full guide: How Privacy Policies Actually Work: A Practical Explanation
What to do and how to stay safe
- Review API documentation for the newly opened Search and Fetch endpoints to understand data handling and rate limits.
- Test the deepseek-flash model endpoint with non-critical workloads to verify output quality and latency.
- Monitor vendor announcements for hardware specifications and weight availability for the V4.1-Flash-1017 checkpoint.
- Verify that your integration with DeepSeek Harness or other agent frameworks is compatible with the recommended v0.2.1-rc.1 version.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What benchmarks did DeepSeek-V4.1-Flash-1017 improve upon?
The model improved on all 11 benchmarks reported by DeepSeek, with gains between 2.0 and 11.1 points compared to the previous version.
Is the model open source?
The September V4.1-Flash weights are MIT-licensed on Hugging Face, but the specific weights and hardware requirements for the 1017 checkpoint are not disclosed.
