DeepSeek puts its whole app behind an API as the open-model serving race heats up

What DeepSeek shipped

DeepSeek has made the capabilities behind the DeepSeek App available through the DeepSeek API, so developers can recreate the full app experience or build custom workflows from standalone pieces. The V4.1-Flash model supports built-in tool calls and handles context assembly server-side, which means the model does the work of gathering context rather than leaving it to the developer's code. Web search supports citations in non-thinking mode, a Deep Thinking Search mode searches and reads pages during multi-step reasoning, and file input parses PDFs, Office files and images into conversation context. m:…4947

Search, Fetch, OCR, speech recognition and text-to-speech are available as standalone APIs, and URL reading automatically fetches links in user messages to inform answers. Safety checks screen inputs and outputs and return refusal indicators when risks are detected, and public-facing applications are required to enable them. Developers remain responsible for content moderation, model registration and labelling AI-generated content, and extended capabilities incur tool usage fees on top of standard token charges. The app experience, in other words, is now a set of billable parts. m:…4947

The model underneath

The model behind the new API is DeepSeek V4.1 Flash, which Verda Inference now serves through an OpenAI-compatible API. Verda describes it as a 552B-parameter Mixture-of-Experts model with a 1M-token context window, priced at $0.15 per million input tokens, with cached input at $0.006 per million and output at $0.60 per million. It activates 8B parameters while reading input and 16B while writing output, and maintains a KV cache of 890 bytes per token. Verda says its deployment, using the SGLang engine, reaches output speeds peaking at approximately 29,300 tokens per second per GPU. Those are the vendor's numbers, and they are the reason the model is interesting to builders: a giant model priced like a small one, if the serving holds up. m:…7528

The catch with models this size is that they do not fit on one machine. NVIDIA's own developer guidance makes the point plainly: the largest models no longer fit on a single machine because frontier-scale models require keeping all weights in GPU memory. Its worked example serves GLM-5.3, a quantized version of Z.ai's mixture-of-experts model, across two DGX Stations connected over ConnectX-8 RDMA, using vLLM tensor parallelism to split weights and computation across GPUs into a single endpoint. That is the DIY version of what Verda is selling as a service. m:…1810

The serving stack is the story

Intel's move this week points the same way. Intel and RadixArk announced a collaboration to optimize the open-source SGLang inference stack for Intel platforms, covering CPUs, GPUs and accelerators, and to upstream the optimizations directly into the core framework rather than keeping them as a separate deployment. The collaboration uses Intel Advanced Matrix Extensions for hardware acceleration and aims to give developers Day-0 readiness for inference software on Intel hardware. The stated goal is to help developers deploy models more efficiently and cost-effectively, with flexible deployment across existing data center and cloud capacity. m:…6301

The pattern across all four announcements is that the model is no longer the hard part to obtain. DeepSeek will sell you the app as an API, Verda will serve the 552B model at small-model prices, NVIDIA will show you how to split it across two boxes, and Intel wants the open serving stack to run well on its hardware. The decision that matters for developers is now which serving path to bet on: a hosted API, a managed inference platform, your own multi-GPU rig, or an open stack tuned to cheaper hardware. Each one trades control for convenience, and the prices quoted this week only hold if the throughput claims survive real workloads. m:…4947 m:…7528 m:…1810 m:…6301