DeepSeek V4.1 Flash Gets New Model With Big Coding Gains, Cheaper Agent Tasks
Chinese AI company DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated version of its Flash model with improvements in coding, agentic tasks, speed, and output efficiency.
The model is available through DeepSeek’s website, app, and API. The company has also published its weights on Hugging Face and ModelScope.
DeepSeek says the update delivers better benchmark performance than its predecessor while completing tasks faster and generating shorter responses. The company describes this as a Pareto improvement in the balance between performance and cost.
New Training Method
For end-to-end software development, DeepSeek uses a post-training method called reinforcement learning with self-feedback, or RLSF.
The technique allows the model to use its own confidence scores or generated critiques as reward signals rather than relying on human evaluators or external reward models.
DeepSeek says this improves the model’s ability to build complete projects from scratch, covering tasks including game development, 3D modeling, and design work.
The model can also create compact animated explainer videos.
DeepSeek tuned V4.1-Flash-1017 for development environments including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. The company recommends DeepSeek Harness version v0.2.1-rc.1 or newer.
Smaller Model Beats Larger Rivals
According to DeepSeek’s own benchmarks, V4.1-Flash-1017 improves over the previous V4.1-Flash across every tested category.
The company also claims the smaller model outperforms larger systems including Claude Opus 5, GPT-5.6 Sol, Kimi K3 and GLM 5.3 in several coding and AI agent benchmarks.
On the DeepSWE benchmark, V4.1-Flash-1017 reached a top score of 77.0% at an average cost of just under $0.20 per task.
At its low setting, the model still scored above 72% while costing around $0.10 per task.
Its predecessor reached 74.2% but cost roughly $0.50 per run.
DeepSeek’s chart places models including GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 at between roughly $2 and more than $10 per task, while showing lower accuracy than V4.1-Flash-1017.
Independent benchmarks are not yet available, and real-world developer assessments of code quality may differ from benchmark results.
Search Engine Opened to Developers
Alongside the new model, DeepSeek is opening its internal web search technology to outside developers for the first time.
The new Search and Fetch APIs use DeepSeek’s own search engine, which the company says already handles hundreds of millions of queries per day worldwide.
Each endpoint costs $0.75 per 1,000 calls.
DeepSeek API users can access the tools for free through server-side integrations in DeepSeek Harness, Claude Code and Codex until October 31, 2026, at 4:00 PM UTC.
China Could Restrict Open-Weight Models
The release comes as the Chinese government reportedly considers tighter controls on publishing open-weight AI models and other key AI technologies.
Under the reported proposal, whether a model ranks among the 10 most capable systems could be one factor in determining whether additional restrictions apply.
That could complicate deeper reliance on Chinese open-weight models if tighter export or publication rules are introduced.
The post DeepSeek V4.1 Flash Gets New Model With Big Coding Gains, Cheaper Agent Tasks appeared first on ProPakistani.
