DeepSeek V4.1 Flash 1017 released with open weights and first search API
Decision Brief
What changedDeepSeek released the V4.1-Flash-1017 model, beating larger competitors on coding and agent tasks at lower per-task cost, and opened its internal search engine to developers for the first time via Search and Fetch APIs.
Why it mattersPer-task cost drops to about $0.20, less than half the previous generation's $0.50, and Claude Code and Codex users can call the search API free until October 31—a direct cost-saving signal for developers building agents on APIs.
Who should careTeams building on model APIs
Affected stackDeepSeekHugging Face
Builder actionWorth evaluating: can DeepSeek replace or complement your current model
Source confidenceMedium · Reliable media or first-hand reporting
DeepSeek-V4.1-Flash-1017 is live on the official site, app, and API, with open weights published on Hugging Face and ModelScope. The model uses reinforcement learning with self-feedback (RLSF) for post-training, using its own confidence or self-critique as the reward signal, eliminating human evaluators or external reward models. Outputs are shorter and tasks finish faster, supporting end-to-end projects from game programming to 3D modeling. It can also be tuned with DeepSeek Harness v0.2.1-rc.1 or later, plus Tencent CodeBuddy and WorkBuddy.
Summary basis: full article readCompiled from the source scope noted above; the original remains authoritative.
Sources
- The Decoder：AI News
- The Decoder：AI News
留言
登入后即可留言，和其他 builder 交换实测心得。
还没有留言，抢头香。