DeepInfer: DeepSeek Open-Sources Its V4.1 Flash Inference Engine
DeepSeek has open-sourced DeepInfer, the inference engine behind its production services, with support for V4.1 Flash. It is an infrastructure release, not a new model. The code is available under Apache 2.0, but this release provides offline inference rather than a complete hosted API. See the official DeepInfer repository.
The release was reported on October 20, 2026. Sources checked October 21; this guide does not report independent performance tests.
What does DeepInfer include?
The official README lists optimized kernels, distributed execution, accelerated decoding and asynchronous scheduling. Supported hardware is NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Its default checkpoint is DeepSeek-V4.1-Flash-1017.
A complete inference service still needs a KV cache system, vision-encoder/prefill/decode disaggregation, and load balancing across replicas. The published benchmark setup uses 4 SM100 GPUs for prefill and 64 for decode, each within one scale-up domain. Those are benchmark configurations, not minimum hardware requirements or a promise of performance on your machine.
What is deepseek-recipe for?
The companion deepseek-recipe project handles the boundary between API requests and model tokens: request conversion, prompt rendering, token encoding and response parsing. It offers Rust libraries and Python bindings, with adapters for Messages, Chat Completions and Responses formats. Its code and public documentation use the MIT license.
The caller still supplies inference, tool execution and HTTP transport. That distinction matters when evaluating a migration: request-format support alone does not mean you have a deployed, authenticated service. Test your actual streaming, tool-call and structured-output workflows before routing users to a replacement backend.
Will self-hosting make DeepSeek cheaper?
Treat lower cost as a hypothesis to test against your workload. Open-source software removes neither accelerator costs nor the work of operating a service. A useful comparison is total monthly serving cost divided by successfully completed tasks. Include idle capacity, retries, monitoring and engineering time, then compare that with your hosted API bill.
For an initial evaluation, use the same representative prompts and output limits on both routes. Measure time to first token, completion latency, failures and task quality at your expected concurrency. Average token throughput alone can hide the slow responses your users experience. This article makes no percentage-savings claim and does not infer an API price change from the code release.
Self-host, use an API, or just chat?
| Your goal | A practical starting point |
|---|---|
| Write, code or ask questions today | Use a hosted chat interface; no infrastructure setup is needed. |
| Add DeepSeek to an application | Start with a hosted API and measure real usage before committing to hardware. |
| Control the serving stack with an infrastructure team | Evaluate DeepInfer against your workload and operational requirements. |
For developers choosing the API route, explore Bumo Router and check its current model catalog and prices. Compare the cost of your expected traffic with self-hosting before deciding. This is a hosted API option; we are not claiming that Router runs the newly released DeepInfer engine.
Try it on aiwith
aiwith's DeepSeek model configuration lists V4.1 Flash (deepseek-flash, the default) and V4 Pro (deepseek-v4-pro). Open DeepSeek chat on aiwith to use the hosted model without setting up an inference server. Availability here refers to the model options, not an installation of DeepInfer; no model-generation benchmark was run for this article.
If your main question is how to access DeepSeek without another login, read our DeepSeek without login guide for access routes and their limits. An open-source engine and account requirements for a hosted website are separate questions.