DeepSeek has released version V4.1 Flash 1017.
Its results on the DeepSeek SWE V1.1 benchmark are kind of unbelievable—the chart on its own benchmark shows it crushing every model!
Its overall capabilities have improved significantly again, while also reducing task execution time and output length, so it can complete the same tasks at a lower cost.
They also emphasized this time that it can develop an entire project from scratch:
No human intervention needed—just very simple prompts, like asking OpenAI 5.5 to make a video or a game in a single sentence.
Another important point is that they’ve opened up their in-house search API.
DeepSeek’s search results for information and data within China are extremely accurate and powerful, so this API should be quite useful. You can hook it up to your Harness.
Its results on the DeepSeek SWE V1.1 benchmark are kind of unbelievable—the chart on its own benchmark shows it crushing every model!
Its overall capabilities have improved significantly again, while also reducing task execution time and output length, so it can complete the same tasks at a lower cost.
They also emphasized this time that it can develop an entire project from scratch:
No human intervention needed—just very simple prompts, like asking OpenAI 5.5 to make a video or a game in a single sentence.
Another important point is that they’ve opened up their in-house search API.
DeepSeek’s search results for information and data within China are extremely accurate and powerful, so this API should be quite useful. You can hook it up to your Harness.