Tarun SureshandClaude Opus 4.6
bf24ffc527
Add Claude Opus-guided optimization loop that jointly optimizes agent configs across multiple benchmarks (TerminalBench-native, GAIA, HLE) with weighted accuracy aggregation and Pareto frontier computation. Key additions: - MultiBenchTrialRunner with native terminal-bench v2 Docker execution - LLM optimizer: fixed params injection, structured trial feedback - OptimizationStore (SQLite) with per-benchmark score persistence - System prompt passthrough from optimizer → eval → agent - Custom terminal-bench agent (OpenJarvisTerminalBenchAgent) avoiding LiteLLM serialization issues with agent_import_path - TOML configs for Qwen3-235B joint agentic optimization - CLI: `jarvis optimize` command with dry-run support Early results on Qwen3-235B-A22B-Instruct-2507-FP8 (4xA100): Best config: native_openhands, temp=0.0, 20 turns → 5.6% weighted acc (TB2=5%, GAIA=6%, HLE=6%) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Programming abstractions for on-device AI.
OpenJarvis is a framework for building AI systems that run entirely on local hardware. Rather than treating intelligence as a cloud service, OpenJarvis provides composable abstractions for local model selection, inference, agentic reasoning, tool use, and learning — all aware of the hardware they run on.
from openjarvis import Jarvis
j = Jarvis() # auto-detect hardware + engine
response = j.ask("Explain backpropagation") # route to best local model
j.ask("Solve x^2 - 5x + 6 = 0", # multi-turn agent with tools
agent="orchestrator",
tools=["calculator", "think"])
j.memory.index("./papers/") # index documents into local storage
results = j.memory.search("attention mechanism") # semantic retrieval
j.close()
Installation
pip install openjarvis # core framework
pip install openjarvis[server] # + FastAPI server
You also need a local inference backend: Ollama, vLLM, SGLang, or llama.cpp.
Quick Start
The fastest path is Ollama on any machine with Python 3.10+:
# 1. Install OpenJarvis
pip install openjarvis
# 2. Detect hardware and generate config
jarvis init
# 3. Install and start Ollama (https://ollama.com)
curl -fsSL https://ollama.com/install.sh | sh
ollama serve # start the Ollama server
# 4. Pull a model
ollama pull qwen3:8b
# 5. Ask a question
jarvis ask "What is the capital of France?"
# 6. Verify your setup
jarvis doctor
jarvis init auto-detects your hardware and recommends the best engine. After init, it prints engine-specific next steps. Run jarvis doctor at any time to diagnose configuration or connectivity issues.
The Five Pillars
| Pillar | What it does | Key abstractions |
|---|---|---|
| Intelligence | Model management and routing | RouterPolicy, QueryAnalyzer, ModelCatalog |
| Engine | Inference runtime abstraction | InferenceEngine ABC — Ollama, vLLM, SGLang, llama.cpp, MLX |
| Agents | Pluggable reasoning strategies | BaseAgent ABC — Simple, Orchestrator, ReAct, OpenHands, OpenClaw |
| Tools | Capabilities via MCP | BaseTool ABC — calculator, code interpreter, web search, memory; external MCP servers auto-discovered |
| Learning | Trace-driven adaptation | LearningPolicy ABC — SFT (model routing), AgentAdvisor (restructuring), ICL (tool usage) |
Every interaction produces a Trace — a structured record of the full reasoning chain. Learning policies consume traces to improve model selection, agent behavior, and tool usage over time.
About
OpenJarvis is part of Intelligence Per Watt, a research initiative studying the efficiency of on-device AI systems. The project is developed at Hazy Research and the Scaling Intelligence Lab at Stanford SAIL.
Sponsors
Laude Institute • Stanford Marlowe • Google Cloud Platform • Lambda Labs