Papers · Nomae Developers

A Programming Paradigm for Spatiotemporal Composability

Yifan Shi, Wei Zhang, and Tianyi Cui · Peking University and DeepSeek-AI

The short version

Dynamic systems should be able to change one component without disturbing everything around it.

The paper gives a formal model for runtime composition: when a component is removed, its tracked changes are safely undone; when its dependencies change, the component reacts in a structured way. The authors implement the model as Cordis, including configuration reconciliation and hot module replacement, and examine it in Koishi's plugin ecosystem.

Read the paper on arXiv

Core distinction

Temporal composability

Can this component leave without leaving its changes behind?

Spatial composability

Can this component declare and adapt to the components it needs?

Preprint · 92 pages · cs.PL / cs.SE

Key points

What the paper contributes.

Make cleanup part of the action

A component effect returns its inverse. The runtime accumulates those inverses and runs them in reverse order when the component unloads, rather than relying on a separate, hand-written uninstall path.

Treat dependencies as live conditions

Components declare what they need. When a provider appears, disappears, or is replaced, only the affected components activate, deactivate, or reload.

Keep a stable view during a transition

A component reads one committed dependency view while it is active or tearing down. That avoids a transition observing a mixture of old and new providers.

Interesting in practice

Useful ideas for agent harnesses and plugin systems.

Give every agent-harness capability an explicit lifecycle: acquire resources through a tracked context and return a disposer beside the setup code.

Represent tools, memory, model access, permissions, and execution environments as declared capabilities. Replace a provider by re-evaluating its direct consumers, not by restarting the whole harness.

Before allowing hot reload, classify each operation: reversible within the system boundary, irreversible but delayable, or irreversible and requiring a compensating action.

Make the dependency graph inspectable. Cycles can be found before activation; often the right fix is to pull the integration into a smaller third component.

Read this as a design foundation, not a benchmark result. The Koishi study shows an adopted implementation across a large plugin ecosystem, but the authors explicitly describe it as observational and leave comparative overhead and developer-productivity measurements for future work.

Added · August 2026

FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution

Shuo Yang et al. · 11 authors · cs.DC

The short version

Local frontier inference is an orchestration problem, not simply a question of whether the model fits in VRAM.

FreeToken is an edge-native serving system for mixture-of-experts models. It coordinates a consumer machine's GPU, CPU, host memory, and PCIe bandwidth to keep agent workloads responsive: it overlaps prefill transfers with computation, shares decode misses between CPU execution and GPU cache fills, and preserves agent state across context edits.

Read the paper on arXiv

Reported range

8 GB laptop GPU

35B Qwen3.6 model at 39.3 decode tokens per second.

32 GB gaming desktop

Interactive serving of a 284B DeepSeek-V4-Flash model.

Preprint · 16 pages · systems evaluation on six machines

Key points

What FreeToken changes.

Schedule against measured bandwidth

For a missed expert, FreeToken chooses how much work to transfer to the GPU and how much to execute on the CPU from the machine's measured host-memory and PCIe bandwidth, rather than using one fixed offloading rule.

Cache the semantics of an agent session

Agent harnesses tend to edit context at tool calls and reasoning boundaries. Checkpointing recurrent state at those boundaries means an edit can recompute only the new suffix instead of re-prefilling a long context.

Treat local resources as elastic

The GPU expert cache can be rebuilt at safe points when available VRAM changes, without restarting the engine or reloading the complete expert pool held in host memory.

Interesting in practice

Useful ideas for local agent runtimes.

Benchmark an agent workload, not only a one-shot prompt. Tool calls, growing context, and repeated prefills expose the latency tails that decide whether a local engine feels usable.

For a hybrid runtime, model transfer, execution, and caching as one scheduling problem. Better prediction alone cannot resolve a cache miss when the underlying bandwidth allocation is wrong.

Use workload semantics as cache boundaries. A tool result or an editable reasoning segment is a more useful checkpoint signal than a uniformly spaced token interval.

Make resource adaptation restart-free. On shared personal hardware, a serving system must tolerate browser, desktop, and other applications changing its usable memory budget at runtime.

Treat the performance numbers as configuration-specific. The paper compares FreeToken with llama.cpp, Ollama, KTransformers, and MoE-Infinity on particular models, quantizations, and machine configurations. The transferable contribution is the adaptive serving design; the exact throughput depends on the model, memory hierarchy, and agent workload.

Papers worthcarrying into the work.