斯坦福大学和 Lambda Labs 的研究人员发表了 OpenJarvis 的研究论文，这是一个完全在设备端运行推理、智能体、记忆和学习的开源框架。
通过 OpenJarvis 配置的开源权重模型，在研究的基准测试协议下，平均性能与最佳云端模型相差不到 3.2 个百分点，而每次查询的边际 API 成本约为云端模型的 1/800，延迟约为云端模型的 1/4。这项研究工作建立在研究团队早先的《每瓦特智能》研究基础之上，该研究报告指出，本地模型已能以交互式延迟处理 88.7% 的单轮对话和推理查询，且从 2023 年到 2025 年，智能效率提升了 5.3 倍。
模型概览与访问
OpenJarvis 并非单一模型。它是一个框架，可将任何受支持的模型与可配置的智能体栈组合使用，并在来自四个模型家族的 11 个本地模型上进行了评估。
| 属性 | 值 |
|---|---|
| 许可证 | Apache 2.0 |
| 框架发布 | 2026 年 3 月 12 日 |
| 论文 | arXiv:2605.17172（发布于 2026 年 5 月 16 日） |
| 代码仓库 | github.com/open-jarvis/OpenJarvis |
| 星标 / 复刻 | ~5.4k / ~1.2k（2026 年 6 月） |
| 编程语言 | Python（约 83%）、Rust（约 9%）、TypeScript（约 7%） |
| 已评估模型 | 来自 4 个模型家族的 11 个本地模型：Qwen3.5、Gemma4、Nemotron、Granite |
| 云端基线模型 | Claude Opus 4.6、GPT-5.4、Gemini 3.1 Pro |
| 支持的推理引擎 | Ollama、vLLM、SGLang、llama.cpp、Apple Foundation Models、Exo（以及其他） |
| 上下文窗口 | 取决于模型 |
| 安装 | 单条命令；宽带环境下约 3 分钟 |
| 硬件 | 已在 7 个平台上测试，从 Mac Mini M4 到 NVIDIA DGX Spark |
架构：五个基本组件与一份配置规范
OpenJarvis 将个人 AI 系统分解为五个类型化的基本组件，通过一个称为配置规范的声明式配置对象进行组合。
- 智能——模型、权重、生成参数和量化格式。
- 引擎——推理运行时（Ollama、vLLM、SGLang 等）、批处理、KV 缓存设置和硬件路径。
- 智能体——推理循环（ReAct 或 CodeAct）、系统提示词、工具使用策略和轮次限制。
- 工具与记忆——外部接口、检索后端、25 个以上数据连接器和 32 个以上消息通道，支持原生 MCP 和可互换的记忆后端。
- 学习——即根据追踪记录更新规格的优化器。该插槽可接受 LoRA、DSPy、GEPA 或大语言模型引导的规格搜索。
每个原语均可独立替换，一份规格会将全部五个原语序列化到一个 TOML 文件中。两份规格可以共享相同的智能体和工具配置，仅模型和引擎不同，因此同一行为无需重写提示词即可在 Mac Mini 和工作站上运行。
大语言模型引导的规格搜索是第二项贡献。这是一种本地-云端协作：前沿云端模型在搜索时充当教师，读取追踪记录、诊断故障集群，并在智能、引擎、智能体和工具与记忆四个维度提出修改建议。仅当修改能改善目标故障集群且不会在其他地方造成显著退化时，该修改才会被接受——研究团队称之为"门控"（默认容忍度 1%）。优化后的规格在推理时完全在设备端运行，零云端调用。教师模型仅在搜索时使用；按每天 100 次查询计算，六个月后分摊的教师成本降至每次查询不到 0.001 美元。
先前的工作（GEPA、DSPy、LoRA）一次只优化一个原语，而仅靠提示词优化器只能弥补云端-本地差距中约 5 个百分点的性能损失。大语言模型引导的规格搜索能弥补 13-32 个百分点，因为它联合跨原语进行修改，优化成本比单原语基线低 7-11 倍。四原语移动空间贡献了 5.5-16.5 个百分点，而大语言模型提议器在相同移动空间下，相比进化搜索平均额外贡献约 10 个百分点。
能力与性能
OpenJarvis 在涵盖 508 个任务的 8 个基准上进行了评估：工具调用（ToolCall-15）、智能体工作流（PinchBench）、编程（LiveCodeBench）、客户服务（τ-Bench V2、τ²-Bench Telecom）、通用助手（GAIA）以及深度研究（LiveResearchBench、DeepResearchBench）。
替换测试：在现有框架（OpenClaw、Hermes Agent）中将预期的云端模型替换为 Qwen3.5-9B，准确率下降 25-39 个百分点。在 OpenJarvis 规格下使用相同模型，残余下降缩小至 5.6-16.5 个百分点——恢复了 56-77% 的可移植性损失。
精度前沿：最佳单机本地模型 Qwen3.5-122B 的平均精度达到 80.3%，而 Claude Opus 4.6 为 83.5%——差距为 3.2 个百分点。在 8 项基准测试中，本地模型在 4 项上达到或超越了云端水平：ToolCall-15、PinchBench、LiveCodeBench 和 τ-Bench V2。
成本与延迟：本地配置构成了精度-效率前沿。Qwen3.5-122B 以每次查询约千分之一美分的成本实现了 80.3% 的精度，而 Claude Opus 4.6 每次查询成本为 0.009 美元——边际 API 成本优势约为 800 倍。在智能体工作负载上，端到端延迟降低了约 4 倍，不过论文指出，单次提示词可能更有利于云端服务。
搜索增益：LLM 引导的规范搜索将 Qwen3.5-9B 学生模型在 PinchBench 上的表现提升至 100%，在 LiveCodeBench 上提升至 83%，在 LiveResearchBench 上提升至 91%。在完整的八项基准测试套件中，每个学生模型的平均增益范围为 13.1 到 31.5 个百分点。作者报告称，这些增益通过了其稳健性检验（奖励权重变体、搜索种子方差和随机重启）。
如何使用
安装只需一条命令。在 macOS、Linux 或 WSL2 上：
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bashWindows 用户运行等效的 PowerShell 脚本（irm … | iex）。安装程序会在宽带环境下约三分钟内配置好 uv、Python 虚拟环境、Ollama 和一个入门模型。桌面 GUI 以 .dmg、.exe、.deb、.rpm 或 .AppImage 格式从发布页面提供。
安装完成后，jarvis 会启动一个聊天会话。入门预设涵盖了常见工作流程：
jarvis init --preset morning-digest-mac # daily briefing with TTS
jarvis init --preset deep-research # multi-hop research with citations
jarvis init --preset code-assistant # agent with code execution and shell access
jarvis init --preset scheduled-monitor # stateful agent on a schedule该框架内置了八个智能体，支持三种执行模式——按需、定时和持续。它连接 25 个以上的数据源（Gmail、日历、iMessage、Notion、Obsidian、Slack、GitHub 等），并通过 32 个以上的消息渠道（WhatsApp、Telegram、Discord、iMessage、Signal 等）暴露智能体。
技能可以从外部目录导入——约 150 个来自 Hermes Agent，约 13,700 个社区技能来自 OpenClaw——所有这些都遵循 agentskills.io 规范。`jarvis optimize skills --policy dspy` 命令可根据本地追踪历史对其进行优化。
Marktechpost 的可视化讲解
OpenJarvis · 斯坦福大学
斯坦福大学 · Hazy Research + Scaling Intelligence Lab
OpenJarvis
一个开源、本地优先的个人 AI 智能体框架，其推理、智能体、记忆和学习功能完全在设备端运行。
与最佳云端方案差距在 3.2 个百分点以内
边际 API 成本降低约 800 倍
延迟降低约 4 倍
Apache 2.0 许可 • arXiv:2605.17172 • 框架于 2026 年 3 月 12 日发布
它是什么
在你的硬件上运行的个人 AI
大多数“个人”AI 仍然将每一次查询都通过云端 API 路由。OpenJarvis 将本地优先作为默认方式，仅在必要时调用云端——这基于该团队在“每瓦特智能”研究中的发现，即本地模型已能处理 88.7% 的单轮查询。
许可协议
Apache 2.0
代码仓库
github.com/open-jarvis/OpenJarvis
模型
11 个本地模型 · 4 个系列
Qwen3.5、Gemma4、Nemotron、Granite
推理引擎
Ollama、vLLM、SGLang、llama.cpp、Apple FM、Exo
架构
五个原语，一份规范
个人 AI 系统被分解为五个类型化、可独立替换的原语，通过一份声明式规范组合而成，该规范序列化为可移植的 TOML 格式。
- 智能——模型、权重、生成参数、量化
- 引擎——推理运行时、批处理、KV 缓存、硬件路径
- 智能体——推理循环（ReAct 或 CodeAct）、提示词、工具策略
- 工具与记忆——25 个以上连接器、32 个以上通道、原生 MCP
- 学习——优化器插槽：LoRA、DSPy、GEPA 或规范搜索
关键方法
大语言模型引导的规范搜索
一个前沿云端模型在搜索时充当教师角色：它读取轨迹、诊断失败集群，并提出跨原语的编辑建议。一个门控机制仅接受不会导致性能倒退的编辑。优化后的规范随后完全在设备端运行——推理时零云端调用。
13 到 32 个百分点
的云端-本地差距被弥合
与单原语基线相比，优化成本更低
四原语移动空间增加了 5.5 到 16.5 个百分点；在相同移动空间下，大语言模型提议者比进化搜索高出约 10 个百分点。
性能
接近云端，成本远低
3.2 个百分点
差距：Qwen3.5-122B 80.3% 对比 Claude Opus 4.6 83.5%
本地模型达到或超越云端模型的基准测试数量
- 在 ToolCall-15、PinchBench、LiveCodeBench、τ-Bench V2 上达到或超越云端水平
- 边际 API 成本降低约 800 倍；延迟降低约 4 倍（论文协议下）
- 交换测试：在规范下，25 到 39 个百分点的下降缩小至 5.6 到 16.5 个百分点（恢复了 56% 到 77%）
开发者体验
从零开始，数分钟内构建一个智能体
一条命令即可配置 uv、Python 虚拟环境、Ollama 以及一个入门模型（宽带环境下约需 3 分钟）：
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bash- 8 个内置智能体，支持按需、定时和持续三种运行模式
- 25 个以上数据连接器 · 32 个以上消息通道
- 通过 agentskills.io 获取技能：来自 Hermes Agent 约 150 个，来自 OpenClaw 约 13,700 个
核心结论
一个研究平台与一个生产基础
OpenJarvis 以约 3.2 个百分点的精度损失——该差距主要集中在推理和研究密集型任务上——换取了显著的成本、延迟和隐私优势。推理、智能体状态和记忆默认全部在设备端运行；云端教师模型为可选且受控。
注意事项：结果取每种配置下 5 次运行的平均值，使用 GPT-5-mini 作为评判模型，且仅在单台机器上运行。采用 Apache 2.0 许可证并持续维护——按作者的说法，这是“秉承 PyTorch 精神”为本地 AI 构建的。
为机器学习工程师解读 AI 研究与开发者工具——
marktechpost.com
关键要点
- OpenJarvis 将推理、智能体、记忆和学习完全在设备端运行，与最佳云端模型的差距控制在 3.2 个百分点以内，而边际 API 成本降低约 800 倍，延迟降低约 4 倍。
- 一种类型化的“规范”将技术栈分解为五个可替换的原语——智能（Intelligence）、引擎（Engine）、智能体（Agents）、工具与记忆（Tools & Memory）以及学习（Learning）——并序列化为可移植的 TOML 格式。
- 大语言模型引导的规范搜索使用前沿云端模型作为搜索时的教师，以 7 至 11 倍更低的优化成本弥补云端与本地之间 13 至 32 个百分点的差距，随后在本地运行且无需任何云端调用。
- 在 8 项基准测试中的 4 项（ToolCall-15、PinchBench、LiveCodeBench、τ-Bench V2）上，本地规范的表现达到或超越了云端；其余差距主要集中在推理和研究密集型任务上。
Researchers at Stanford University and Lambda Labs, have published the research paper for OpenJarvis, an open-source framework that runs inference, agents, memory, and learning entirely on-device.
The open-weight models configured through OpenJarvis land within 3.2 percentage points of the best cloud model on average, at roughly 800× lower marginal API cost per query and roughly 4× lower latency under the research’s benchmark protocol. This research work builds on the research team’s earlier Intelligence Per Watt study, which reported that local models already handle 88.7% of single-turn chat and reasoning queries at interactive latency, with intelligence efficiency improving 5.3× from 2023 to 2025.
Model Overview & Access
OpenJarvis is not a single model. It is a framework that composes any supported model with a configurable agent stack, evaluated across 11 local models from four families.
| Property | Value |
|---|---|
| License | Apache 2.0 |
| Framework release | March 12, 2026 |
| Paper | arXiv:2605.17172 (posted May 16, 2026) |
| Repository | github.com/open-jarvis/OpenJarvis |
| Stars / forks | ~5.4k / ~1.2k (June 2026) |
| Languages | Python (~83%), Rust (~9%), TypeScript (~7%) |
| Evaluated models | 11 local models across 4 families: Qwen3.5, Gemma4, Nemotron, Granite |
| Cloud baselines | Claude Opus 4.6, GPT-5.4, Gemini 3.1 Pro |
| Supported engines | Ollama, vLLM, SGLang, llama.cpp, Apple Foundation Models, Exo (among others) |
| Context window | Model-dependent |
| Installation | Single command; ~3 minutes on broadband |
| Hardware | Tested on 7 platforms, from Mac Mini M4 to NVIDIA DGX Spark |
Architecture: Five Primitives and a Spec
OpenJarvis decomposes a personal AI system into five typed primitives, composed through a single declarative configuration object called a spec.
- Intelligence — the model, weights, generation parameters, and quantization format.
- Engine — the inference runtime (Ollama, vLLM, SGLang, etc.), batching, KV-cache settings, and hardware path.
- Agents — the reasoning loop (ReAct or CodeAct), system prompts, tool-use policy, and turn limits.
- Tools & Memory — external interfaces, retrieval backends, 25+ data connectors, and 32+ messaging channels, with native MCP support and interchangeable memory backends.
- Learning — the optimizer that updates the spec from traces. This slot accepts LoRA, DSPy, GEPA, or LLM-guided spec search.
Each primitive is independently swappable, and a spec serializes all five into a TOML file. Two specs can share the same agent and tool configuration and differ only in model and engine, so the same behavior runs on a Mac Mini and a workstation without rewriting prompts.
LLM-guided spec search is the second contribution. It is a local–cloud collaboration: a frontier cloud model acts as a teacher at search time, reading traces, diagnosing failure clusters, and proposing edits across Intelligence, Engine, Agents, and Tools & Memory. An edit is accepted only if it improves the target failure cluster without causing meaningful regressions elsewhere — the research team calls this the gate (default tolerance 1%). The optimized spec then runs entirely on-device at inference time, with zero cloud calls. The teacher is used only at search time; at 100 queries per day, the amortized teacher cost falls below $0.001 per query within six months.
Prior work (GEPA, DSPy, LoRA) optimizes one primitive at a time, and prompt optimizers alone recover only about 5 pp of the cloud–local gap. LLM-guided spec search recovers 13–32 pp because it edits across primitives jointly, at 7–11× lower optimization cost than single-primitive baselines. The four-primitive move space contributes 5.5–16.5 pp, and the LLM proposer adds about 10 pp on average over an evolutionary search at the same move space.
Capabilities & Performance
OpenJarvis was evaluated across 8 benchmarks spanning 508 tasks: tool calling (ToolCall-15), agentic workflows (PinchBench), coding (LiveCodeBench), customer service (τ-Bench V2, τ²-Bench Telecom), general assistance (GAIA), and deep research (LiveResearchBench, DeepResearchBench).
The swap test: Replacing the intended cloud model with Qwen3.5-9B in existing frameworks (OpenClaw, Hermes Agent) drops accuracy by 25–39 pp. With the same model under an OpenJarvis spec, the residual drop shrinks to 5.6–16.5 pp — recovering 56–77% of the portability loss.
The accuracy frontier: The best single local model, Qwen3.5-122B, reaches 80.3% average accuracy versus Claude Opus 4.6 at 83.5% — a 3.2 pp gap. Local specs match or exceed cloud on 4 of 8 benchmarks: ToolCall-15, PinchBench, LiveCodeBench, and τ-Bench V2.
Cost and latency: Local configurations form the accuracy–efficiency frontier. Qwen3.5-122B delivers its 80.3% at roughly a thousandth of a cent per query, versus $0.009 per query for Claude Opus 4.6 — an approximately 800× marginal API-cost advantage. End-to-end latency drops by roughly 4× on the agentic workloads, though the paper notes single-shot prompts can favor cloud serving.
Search gains: LLM-guided spec search improves the Qwen3.5-9B student to 100% on PinchBench, 83% on LiveCodeBench, and 91% on LiveResearchBench. Across the full eight-benchmark suite, average gains per student model range from 13.1 to 31.5 pp. The authors report that these gains survive their robustness checks (reward-weight variants, search-seed variance, and random restarts).
How to Use it
Installation is one command. On macOS, Linux, or WSL2:
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bashWindows users run an equivalent PowerShell script (irm … | iex). The installer provisions uv, a Python virtual environment, Ollama, and a starter model in about three minutes on broadband. A desktop GUI ships as a .dmg, .exe, .deb, .rpm, or .AppImage from the releases page.
After install, jarvis starts a chat session. Starter presets cover common workflows:
jarvis init --preset morning-digest-mac # daily briefing with TTS
jarvis init --preset deep-research # multi-hop research with citations
jarvis init --preset code-assistant # agent with code execution and shell access
jarvis init --preset scheduled-monitor # stateful agent on a scheduleThe framework ships with eight built-in agents across three execution modes — on-demand, scheduled, and continuous. It connects to 25+ data sources (Gmail, Calendar, iMessage, Notion, Obsidian, Slack, GitHub, and others) and exposes agents over 32+ messaging channels (WhatsApp, Telegram, Discord, iMessage, Signal, and others).
Skills can be imported from external catalogs — about 150 from Hermes Agent and about 13,700 community skills from OpenClaw — all following the agentskills.io specification. A jarvis optimize skills --policy dspy command refines them from local trace history.
Marktechpost’s Visual Explainer
OpenJarvis · Stanford
Stanford · Hazy Research + Scaling Intelligence Lab
OpenJarvis
An open-source, local-first framework for personal AI agents that run inference, agents, memory, and learning entirely on-device.
Within 3.2 pp of best cloud
~800× lower marginal API cost
~4× lower latency
Apache 2.0 • arXiv:2605.17172 • Framework released March 12, 2026
What it is
Personal AI that runs on your hardware
Most “personal” AI still routes every query through a cloud API. OpenJarvis makes local-first the default and calls the cloud only when needed — building on the team’s Intelligence Per Watt finding that local models already handle 88.7% of single-turn queries.
License
Apache 2.0
Repository
github.com/open-jarvis/OpenJarvis
Models
11 local models · 4 families
Qwen3.5, Gemma4, Nemotron, Granite
Engines
Ollama, vLLM, SGLang, llama.cpp, Apple FM, Exo
Architecture
Five primitives, one spec
A personal AI system is decomposed into five typed, independently swappable primitives, composed through a single declarative spec serialized to portable TOML.
- Intelligence — model, weights, generation params, quantization
- Engine — inference runtime, batching, KV-cache, hardware path
- Agents — reasoning loop (ReAct or CodeAct), prompts, tool policy
- Tools & Memory — 25+ connectors, 32+ channels, native MCP
- Learning — optimizer slot: LoRA, DSPy, GEPA, or spec search
Key method
LLM-guided spec search
A frontier cloud model acts as a teacher at search time: it reads traces, diagnoses failure clusters, and proposes edits across primitives. A gate accepts only non-regressing edits. The optimized spec then runs entirely on-device — zero cloud calls at inference time.
13–32 pp
of the cloud–local gap closed
lower optimization cost vs single-primitive baselines
The four-primitive move space adds 5.5–16.5 pp; the LLM proposer adds ~10 pp over evolutionary search at the same move space.
Performance
Close to cloud, far cheaper
3.2 pp
gap: Qwen3.5-122B 80.3% vs Claude Opus 4.6 83.5%
benchmarks where local matches or beats cloud
- Matches/exceeds cloud on ToolCall-15, PinchBench, LiveCodeBench, τ-Bench V2
- ~800× lower marginal API cost; ~4× lower latency (paper’s protocol)
- Swap test: a 25–39 pp drop shrinks to 5.6–16.5 pp under a spec (56–77% recovered)
Developer experience
From zero to an agent in minutes
One command provisions uv, a Python virtual environment, Ollama, and a starter model (~3 minutes on broadband):
curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bash- 8 built-in agents across on-demand, scheduled, and continuous modes
- 25+ data connectors · 32+ messaging channels
- Skills via agentskills.io: ~150 from Hermes Agent, ~13,700 from OpenClaw
The bottom line
A research platform and a production foundation
OpenJarvis trades roughly 3.2 pp of accuracy — the gap concentrating on reasoning- and research-heavy tasks — for major cost, latency, and privacy gains. Inference, agent state, and memory stay on-device by construction; the cloud teacher is optional and bounded.
Caveats: results average 5 runs per configuration, use GPT-5-mini as judge, and were run on a single machine. Apache 2.0 and actively maintained — built, in the authors’ words, “in the spirit of PyTorch” for local AI.
AI research and developer tools, decoded for ML engineers —
marktechpost.com
Key Takeaways
- OpenJarvis runs inference, agents, memory, and learning fully on-device, landing within 3.2 pp of the best cloud model at ~800× lower marginal API cost and ~4× lower latency.
- A typed "spec" decomposes the stack into five swappable primitives — Intelligence, Engine, Agents, Tools & Memory, and Learning — serialized to portable TOML.
- LLM-guided spec search uses a frontier cloud model as a search-time teacher to recover 13–32 pp of the cloud–local gap at 7–11× lower optimization cost, then runs locally with zero cloud calls.
- Local specs match or exceed cloud on 4 of 8 benchmarks (ToolCall-15, PinchBench, LiveCodeBench, τ-Bench V2); the remaining gap concentrates on reasoning- and research-heavy tasks.