DeepSeek goes open: what can you actually build with its ecosystem?

By Turing Post · October 31, 2026.

Turing Post notes that DeepSeek has open-sourced far more than a model: weights, an inference engine, an agent harness and the underlying libraries. Taken together, according to the newsletter, they show how an agent system can be built. The central question is how much of that system can actually be used and what it takes to run it. The piece distinguishes between open components and a service that works end to end.

The starting point is that once a capable model is used as an agent, other questions arise beyond what it can do: who can run it, how much it costs and how it connects to tools. The newsletter points out that DeepSeek has released an inference engine, an agent harness and libraries that go all the way down to data movement between chips, so much of the machinery surrounding the model is visible. It also asks what still depends on DeepSeek's hosted services or its private infrastructure.

The article's table of contents includes: the DeepSeek-V4.1-Flash model, DeepInfer, what DeepSeek Harness lets you change, why a protocol library deserves attention, DeepSeek Elastic Compute as a place where agents can run, what still goes through the DeepSeek API, what lies beneath the engine, where the gap is when putting it all together, and what DeepSeek has put in our hands. Only the first two sections and the start of the third are readable in the email; the rest is behind the (Premium) paywall.

The model: DeepSeek-V4.1-Flash. According to the newsletter, it arrived on September 10, 2026 with multimodal reasoning, coding and agentic workflows. DeepSeek says the base model performs comparably to V4-Pro-Base on several benchmarks, doing better on some tasks and worse on others. The V4.1-Flash-1017 update arrived on October 17, and the updated weights are available under an MIT license.

One relevant figure for agents is that the model's global KV cache takes up 890 bytes per token, roughly a quarter of V4 Flash's. A smaller cache can make it easier to maintain long contexts over a session. The newsletter itself clarifies that this measures the global KV state, not the memory footprint of the entire running model.

Architecture: V4.1 Flash is a multimodal mixture-of-experts (MoE) model. Its 552 billion backbone parameters are combined with 196 billion Engram parameters, a sparse lookup memory bank learned during training. Engram gives the model conditional access to externalized knowledge without inflating the active KV cache, and it is distinct from the user's prompt or conversation history.

Its 40 layers are split into a 20-layer causal encoder and a 20-layer decoder. The decoder's global KV state is projected from the encoder's final states, so most prompt tokens can skip the full decoder computation. It uses 8 billion activated parameters per token in prefill and 16 billion in decoding, although local sliding-window attention still requires additional work.

V4.1 Flash also uses "bounded replay": it approximately reconstructs the recent sliding-window state from a short suffix. Some recomputation is traded for lower persistent storage requirements. The newsletter warns that approximation and sparse-selection errors can cause problems in cases not covered by its evaluations.

Compressed Sparse Attention 2 tackles another source of duplication: some layers create global KV state and select relevant entries, while others reuse the state, or both the state and the selection. Combined with a low-precision FP4 cache, this shrinks the global KV footprint. For agents, these choices aim to cut the cumulative cost of long context, although whether it pays off depends on how the serving infrastructure around the model is built.

DeepInfer: according to DeepSeek, this is the inference framework behind its production services, and the public version currently supports offline inference. It includes DSpark speculative decoding: a draft model proposes several tokens, and confidence estimates help decide how many are worth verifying; the speedup depends partly on how many draft tokens are accepted. Asynchronous scheduling keeps the accelerator busy while the CPU prepares the next batch. Supported hardware is NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.

With multiple concurrent users, more unknowns appear: where to send each request and whether already-processed context can be reused. The public version leaves it to developers to set up shared KV cache storage, separate vision/prefill/decode serving and load balancing. The engine manages its internal KV state, but sharing and reusing it across a service requires more infrastructure.

DeepSeek's deployment guide gives an idea of the intended scale: 64 GPUs for prefill plus 64 for decoding is the recommended minimum efficient configuration, although the guide also says a single prefill instance and a single decoding instance can be deployed. The interim conclusion is that DeepInfer lets you inspect and adapt the engine, but serving a coding agent to users requires a serving system around it. The newsletter links this to the next question, how the agent works with tools once the model is running, which is covered in the section on DeepSeek Harness, already behind the paywall.

The email ends with an invitation to subscribe to Premium to read the rest and get a PDF on the DeepSeek ecosystem, and links to a YouTube video about a model that learns to read without a tokenizer. The rest of the article (Harness, the protocol library, Elastic Compute, the API and the lower layers) does not appear in the email, so it is not summarized here.