DeepSeek 4.1 Released with 256-Expert MoE and Open Weights

On October 4, 2026, DEV Community reported that DeepSeek had opened up the weights of DeepSeek 4.1 and made a corporate endpoint available worldwide. The same write-up refers to the release as part of the DeepSeek-V4.1 Foundation series. According to the report, the model uses a 256-expert MoE architecture and adds MTP-4x alongside DualPipe 2.0, a combination the article credits with redrawing the cost boundary for inference at large scale.

That is where the verifiable detail ends. The report lists no parameter count, no context length, no pricing, and no benchmark scores, and it does not say which regions or which customers can reach the corporate endpoint. DeepSeek itself does not appear anywhere in the material: there is no official blog post, model card, or pricing page to check against. As a result, the claim that the weights and the endpoint are now globally available rests on a single outlet, and questions about how the release performs or what it costs remain open here.

One day earlier, on October 3, 2026, TechCrunch published a roundup of AI agents that run inside text messages. The list covers general-purpose assistants as well as agents designed for families, travel, and work, and it is organized by use case rather than by vendor. That article is unrelated to DeepSeek 4.1; it is simply the earlier item in the feed and says nothing about the release's timing, contents, or availability.