DeepSeek’s coding harness says, “Everything is a plugin,” but the more revealing engineering decision was to embed a mature plugin kernel, pin its exact commit, and document 18 changes.
The kernel is Cordis, an MIT-licensed component extracted from the Koishi chatbot framework. It provides the lifecycle machinery for contexts, service injection, typed events, reactive dependencies, and reversible effects. A plugin can be unloaded with its listeners, services, and other side effects without ending the session.
DeepSeek borrowed that plumbing, then modified it. Its changes include repairs for reentrant disposal and a port of an upstream pull request before Cordis merged it. This let DeepSeek focus its product work higher in the stack, around the agent loop, tools, and context.
Borrow the plumbing, differentiate above it
The harness spans 219 packages and includes bridges for Claude Code hooks and Codex configuration. With a borrowed kernel, an open plugin surface, and competitor compatibility, the kernel itself is unlikely to be the main differentiator. Execution and ecosystem have to carry more of the product.
There is a trade-off. Sophisticated lifecycle management adds complexity for users who never need hot swapping. Reversible effects and clean unloading matter when components must change during a session. They matter less in a simple, static system.
Specialized agents need controlled context
Hermes lets each named bot have its own model, provider, tools, skills, chats, and schedules, including local models. Grockbot offers less configuration but clearer agent delegation, isolated virtual computers, and mobile access.
In one Hermes test, delegated research returned, but the assigned bot did not visibly know it had received the work. Delegation happened, yet the product did not make that state clear.
The strongest case for specialized bots is not that naming more agents makes them reason better. It is context control. A specialized bot can receive fewer irrelevant tools, less unrelated history, and narrower instructions. Grockbot addresses the same problem differently by separating agents from personal accounts through isolated virtual computers.
In both designs, the useful capability sits around the model: control over context, tools, identity, and environment.
Memory movement sets the practical limit
Memory can dominate the economics of local AI. Framework’s listed 128 GB system price rose from $1,999 to $3,449, while TrendForce expects a structural DRAM shortage into 2028.
The constraint becomes concrete when tuning larger models. Full AdamW mixed-precision training of a 27 billion parameter model requires 486 GB before examples. Unsloth’s generic 4-bit QLoRA table gives a 22 GB minimum. That can fit narrowly on a 24 GB desktop card, but not on a 16 GB laptop GPU.
Soup demonstrated an 8 billion parameter tune at 3.32 GB peak memory by streaming one decoder layer at a time. Its earlier 4 GB headline predates a correctness repair, and there is still no measured 27 billion parameter run on a small card.
The proposed 27 billion parameter setup also includes untested elements such as linear attention, an extra prediction head, a vision component, adapter targets, and kernels. The conservative recommendation is a 14 billion parameter model on a 16 GB laptop. For 27 billion parameters, use a 24 GB desktop card or rented hardware.
Model size alone is incomplete. The actual constraint is how parameters, optimizer state, examples, caches, and intermediate data move through available memory.
Selective computation needs workload evidence
QWin 3.8 lists 512 experts, with 10 active per token, four key-value heads for 64 query heads, and rope_theta 10 million. These choices affect active compute, context behavior, and concurrent serving capacity. Total parameters do not equal the work performed for every token, while reduced key-value storage can change how many requests a serving system handles concurrently.
The broader pattern is selective computation combined with less memory movement. vLLM reported two to four times higher throughput at the same latency with PagedAttention. DeepSeek reported a 93.3% cache reduction for multi-head latent attention against its dense 67 billion parameter baseline.
These are substantial reported improvements, not universal guarantees. Architectural inspection, vendor reporting, and limited demonstrations do not establish how a complete system will behave under another workload.
Judge the system around the model: how it controls lifecycle, context, memory, serving, and isolation under the workload you actually need.
A Google pricing brief argues that cost and distribution, rather than intelligence, are becoming the moat. Its benchmark figures are too ambiguous for reliable model ranking. The narrower conclusion is better supported: low price matters, but architecture and distribution determine whether cheap intelligence becomes dependable.
Cheap intelligence is only an input. The product is built, or lost, in the infrastructure that makes it usable and dependable.