Modu AI Harness

高性能 · 高并发 · 低内存

High performance · High concurrency · Low memory

AGI时代的Harness底座A general-purpose harness foundation for the AGI era

轻盈启动，从容运行。
Start light. Run at scale.了解核心能力 ↓查看完整界面 ↗

High performance

高性能

让启动更快，让资源回到任务本身。

以 Rust 构建轻量执行底座。所提供的 v0.1.0 基准中，启动时间为 0.011 秒，发布体积为 3.87 MB，降低反复启动与分发部署的开销。

A lightweight Rust foundation. The supplied v0.1.0 benchmark reports a 0.011-second startup and a 3.87 MB release, reducing startup and distribution overhead.

0.011 s启动 / Startup3.87 MB发布体积 / Release size

High concurrency

高并发

更多会话，共享有限的机器资源。

面向多任务与子代理协作的运行场景。所提供的 100 并发实测内存为 0.59 GiB，让并行会话拥有更低的基础资源占用。

Built for multi-task and subagent workflows. The supplied 100-session measurement reports 0.59 GiB of memory, keeping the baseline cost of parallel sessions low.

100并发会话 / Concurrent sessions0.59 GiB内存实测 / Measured memory

Low memory

低内存

每一个会话，都保持轻量。

所提供的单会话内存为 5.97 MiB。更小的常驻开销，为工具调用、上下文处理与多实例运行留出资源空间。

The supplied single-session result is 5.97 MiB. A smaller baseline footprint leaves more room for tool execution, context processing and additional instances.

5.97 MiB单会话内存 / Single-session memory

Cache-aware by design

缓存优化

重点适配 KV cache，让重复上下文更有价值。

围绕稳定的提示词前缀、可复用的上下文组织与连续会话设计优化路径，减少无必要的前缀变化，为支持缓存的模型服务创造更好的命中条件。

Focused on KV cache compatibility: stable prompt prefixes, reusable context structure and session continuity help cache-capable model services reuse repeated context.

稳定前缀系统指令与工具定义保持有序，动态内容后置。
Stable prefixes, dynamic content last.
上下文复用围绕连续任务减少重复上下文的无效改写。
Reuse context across continuing tasks.
KV cache 适配结合模型提供方的缓存机制评估命中、时延与 token 成本。
Evaluate cache hits, latency and token costs by provider.

缓存效果取决于模型服务、前缀一致性与缓存有效期；本页 benchmark 未包含缓存命中率测试。Cache benefits depend on provider support, prefix consistency and cache lifetime. Cache hit rates are not measured in the benchmark below.

Benchmark

从启动、体积到会话内存，逐项查看。Startup, release size and session memory, side by side.

以下根据提供的 v0.1.0 基准原表制作中英双语表格，数值保持原样。100 并发中 modu、jcode 为实测，其余带“推算”标记的数据为估算。测试环境、负载与脚注未随原表提供；启动耗时不代表模型推理速度。Bilingual tables transcribed from the supplied v0.1.0 benchmark, with original values preserved. The 100-session values for modu and jcode are measured; rows marked 推算 are estimates. Environment, workload and footnotes were not supplied. Startup time is not model inference speed.

性能与资源对比 Performance & resource comparison

性能与资源对比 / Performance and resource comparison
#Rank	AgentAgent	技术栈Stack	启动（s）Startup (s)	发布体积Release size	单会话（MiB）Per session (MiB)	100 并发100 concurrent sessions
1	modu 0.1.0	Rust	0.011	3.87 MB	5.97	0.59 GiB实测Measured
2	jcode 0.84.0	Rust	0.038	144.0 MB	32.3	3.14 GiB实测Measured
3	goose 1.45.0	Rust	0.065 ᶜ	250.3 MB	175.3	15.4 GiB推算Estimated
4	codex 0.155.1	Rust	0.120	218.2 MB	312.6	31.2 GiB推算Estimated
5	grok 1.0.34	Rust	0.192 ᵍ	136.4 MB	177.8	13.2 GiB推算Estimated
6	claude 2.1.278	Bun/JS	0.345	207.6 MB	327.1	23.8 GiB推算Estimated
7	dsh 0.1.5-rc.2	TypeScript	0.482	288 MB ⁿ	187.6	13.7 GiB推算Estimated
8	qwen 0.24.1	TypeScript	1.047	119.3 MB	478.6	22.3 GiB推算Estimated
9	opencode 1.18.31	TS on Bun	1.381	137.8 MB	278.2	39.2 GiB推算Estimated
10	pi 0.85.1	TypeScript	1.800 ᵖ	20.9 MB	225.4	12.1 GiB推算Estimated

查看原始表格 / View source table ↗

小屏可横向滑动表格 / Scroll horizontally on smaller screens

相对 modu 的倍数 Relative to modu

相对 modu 的倍数 / Relative to modu
序号#	AgentAgent	启动Startup	体积Release size	单会话Per session	100 并发100 concurrent sessions
1	modu	1.0×	1.0×	1.0×	1.0×
2	jcode	3.5×	37.2×	5.4×	5.4×
3	goose	6.1×	64.7×	29.4×	26.2×
4	codex	11.2×	56.4×	52.4×	53.3×
5	grok	17.9×	35.2×	29.8×	22.5×
6	claude	32.3×	53.6×	54.8×	40.6×
7	dsh	45.0×	74.4×	31.4×	23.4×
8	qwen	97.8×	30.8×	80.2×	38.1×
9	opencode	129.0×	35.6×	46.6×	66.9×
10	pi	168.2×	5.4×	37.8×	20.6×

查看原始表格 / View source table ↗

数值越小表示该项耗时或资源占用越少；沿用原表精度。
Lower values indicate less time or resource usage. Original precision retained.