DeepSeek V4.1 Flash October Checkpoint Adds Search API and Multimodal Input
DeepSeek shipped the V4.1-Flash-1017 checkpoint on October 17, adding native multimodal input and a Search API to V4.1 Flash. The update gives an open-weight model that already tops LiveBench agentic coding the ability to see images and query the web without external tooling scaffolds.
What Changed in the October 17 Checkpoint
The V4.1-Flash-1017 build reports:
- DeepSWE: 77.0 on the new checkpoint
- Multimodal input: native image processing integrated into the V4.1 Flash base
- Search API: built-in web retrieval, the first in the V4.1 Flash line
The September 2026 LiveBench run had V4.1 Flash at 77.3 on agentic coding — 11 points ahead of Fable 5.1 (66.1) and 20 points above GPT-6 Astra (57.3), at $0.029 per task versus Fable 5.1’s $5.51 per task.
Why the Search API Matters for Agentic Use
Agentic tasks regularly require pulling external information mid-task. Without native retrieval, agents depend on tool-calling scaffolds or dedicated retrieval layers, adding latency and failure points. A native Search API at V4.1 Flash’s cost profile changes the economics for developers building research-intensive or information-retrieval-heavy agents.
DeepSeek has distributed prior Flash versions as open weights under MIT licence, making the capability available for self-hosted deployments at no licensing cost.
Benchmark Context
| Benchmark | Score |
|---|---|
| LiveBench Agentic Coding (Sep 2026) | 77.3 |
| LiveBench Overall (Sep 2026) | 81.1 |
| DeepSWE (V4.1-Flash-1017) | 77.0 |
The DeepSWE 77.0 figure is consistent with the model’s prior agentic coding profile. Independent SWE-bench Verified testing on the V4.1-Flash-1017 checkpoint is pending.
DeepSeek’s V4 Flash model shipped in April 2026 with 79% SWE-bench Verified on 13 billion active parameters; V4.1 Flash extended that base with longer reinforcement learning training. The October 17 checkpoint adds multimodal and search capability to the same architecture rather than retraining from scratch.