DeepSeek Gave Away the Engine and Kept the Keys

Share

This week DeepSeek did something frontier labs keep promising but almost never do. It open sourced the actual engine it uses to serve its own models. DeepInfer, the production inference framework running its cheapest and fastest model, V4.1 Flash, is now Apache 2.0 on GitHub. This is not another weights dump or a research paper. It is the serving software itself, down to the optimized kernels and the scheduler. On its face it looks like the most useful gift a frontier lab has handed out this year, and plenty of coverage treats it as pure philanthropy aimed at OpenAI and Anthropic. I read it differently. That openly licensed code is hiding a quiet lock in.

The parts DeepSeek actually opened

Start with what shipped, because it is real and it is good. DeepInfer includes optimized kernels, expert parallelism, speculative decoding under the name DSpark, recovery for a sliding window cache, asynchronous scheduling, and deterministic modes. DeepSeek claims its decode path tops 25,000 tokens per second per GPU across 64 GPUs, and that a single DeepInfer cluster handles 155 trillion tokens a day. I am reporting those as DeepSeek's own numbers, not independently verified. Even taking the claims with salt, the direction is unmistakable. The company that rattled the frontier last year with cheap weights is now competing on the serving layer too.

A companion library called the deepseek recipe ships under MIT. Its whole job is to make the engine speak the exact same API protocol, Chat Completions, Messages, and Responses, as the paid DeepSeek service. That detail matters and I will come back to it.

The catch is not hidden, it is the point

Here is the part most of the headlines glide past. DeepInfer supports exactly one model, DeepSeek V4.1 Flash, and exactly two kinds of silicon, NVIDIA's SM100 datacenter chip and Huawei's Ascend 950. Not Hopper, not a consumer card, not the kind of hardware most founders or even most AI companies actually own. I checked my own rig against the requirement. My workstation runs a single RTX PRO 6000 Blackwell with 96 gigabytes of VRAM, and my agent hosts are dual DGX Sparks. None of that is SM100 class, so no version of me runs DeepInfer today. To use their gift I would have to rent a different chip family entirely.

Then there is what the release leaves out. There is no KV cache store, no prefill and decode separation, no load balancer. Those are precisely the pieces that turn raw kernels into a production service that can serve many users at once. An engineer who pulls DeepInfer down does not get something that resembles the DeepSeek API. They get the computational core and an instruction sheet telling them to build the surrounding infrastructure themselves.

I kept circling the same image. The chocolate is on the table. The wrapper is welded to their brand and their silicon. This is open source the way a free trial is open.

Openness is a strategy, not a promise

I have spent most of this year arguing that openness in AI is frequently a distribution move, and DeepInfer is the cleanest example yet. Look at what this release actually accomplishes for DeepSeek. It does not weaken their lock in, it thickens it. Every developer who wires DeepInfer into a product has made it harder to leave DeepSeek, not easier, because the engine is tuned for DeepSeek's own weights and the integration speaks DeepSeek's own protocol. DeepSeek gives away the engine, and what it gets back is a widening moat around its model, its API, and its datacenter partners, all of it happily subsidized by the open source community's enthusiasm.

The same logic explains why more vendors will copy this play. If you open the software but keep the silicon, the model, and the missing orchestration layers, you extract the goodwill of open source without giving up the economic control. NVIDIA has been running this exact playbook with Dynamo and with its own open models. DeepInfer is proof that the strategy has reached the frontier labs themselves.

The distinction founders keep missing

For a founder trying to build on AI, this is the distinction that matters more than any benchmark table. Open weights is not an open market, and open code is not an open standard. A vendor can hand you source for free and still own the relationship, the compatibility layer, and the upgrade path. The question is not whether the license says Apache 2.0. The question is what you have to give up, or rent, or route through their stack in order to actually use it. The deepseek recipe is the tell. It exists so that anything you build now speaks DeepSeek's protocol, which is a very gentle way of training your entire architecture to depend on the vocabulary of one vendor.

None of this makes DeepInfer bad. It makes it a tool with a bill attached, and the bill is denominated in commitment. If you are already all in on DeepSeek models at datacenter scale, this could be a genuinely efficient way to serve them, and the price signal it sends to Western providers is real. Cheap open models keep dragging the market price of intelligence down, and DeepSeek now competes on the serving layer as well.

What this should change about your stack

The lesson I keep landing on is the same one whether the gift comes from a frontier lab or a chipmaker. The moat was never the model and never the engine. It is the workflow, the data, the distribution, and the trust that live outside whoever ships the software. The engine is a rented asset no matter which silo it comes from. Your process is the only part that cannot be open sourced away from you, and that is the thing worth owning.

That is why I stay deliberately agnostic about engines and about models. I would rather run open software that serves many models on hardware I own, like vLLM and SGLang on my own GPUs, than bet my product on one vendor's tuned stack, however generous the license. General purpose open engines keep every door open. A strategically open engine closes all the doors except the one that leads back to its owner.

The test I would apply

Before you treat any open source AI release as insurance against lock in, ask one question. Could I walk away from this vendor's stack tomorrow and keep running the same product with different software, on hardware I already own? If the answer requires renting their chip class, adopting their model line, or rebuilding the orchestration they deliberately left out, then the openness is serving them, not you. DeepSeek has shipped something genuinely impressive here. Just do not mistake a generous license for a durable escape hatch.