TechPrufer
Get started
Lambda
Lambda logoLambda · Feature trail

Open Source AI Model Harness

2 posts · Jul 2026 – Sep 2026

What this trail covers

The platform now supports NCode, an open-source AI model harness that provides transparency and control over how models operate, what tools they can access, and how their performance is measured. This allows for greater reliability and customization by enabling users to integrate their own models, fine-tune them for specific tasks, and monitor their execution. NCode addresses the 'black-box' problem of proprietary harnesses by offering a fully inspectable and modifiable infrastructure layer, facilitating the retargeting of agentic systems to local models. This post introduces Open Jarvis, a framework that decouples the agent harness from the specific LLM, enabling local models to achieve performance parity with cloud models at significantly lower cost and latency. It details a five-primitive spec (Intelligence, Engine, Agent Logic, Tools & Memory, Learning) that can be optimized for any model and hardware, using a frontier model for guided spec search and improvement. The post also provides concrete performance metrics and cost comparisons for running local models on Lambda's infrastructure.

All Lambda trails
  1. Open Jarvis: making local LLMs work as agentsThis post introduces Open Jarvis, a new framework for local LLM agents that decouples the harness from the model. It details a five-primitive configurable spec (Intelligence, Engine, Agent Logic, Tools & Memory, Learning) and demonstrates how LLM-guided spec search can retune harnesses for local models, recovering performance lost when swapping from cloud to local LLMs. It provides concrete benchmarks and cost comparisons for running local models on Lambda's HGX B200 instances, showing significant cost and latency advantages over cloud models.
  2. OpenResearcher: a reproducible and scalable pipeline for training deep research agentsThis post introduces OpenResearcher, a reproducible and scalable pipeline for training deep research agents. The pipeline consists of three stages: 1) Building a reproducible research environment with 15 million documents, embedding them using Qwen3-Embedding-8B on A100 GPUs, and serving retrieval on H100 GPUs. 2) Generating over 97,000 long-horizon research trajectories using GPT-OSS-120B on 64 H100 GPUs, capturing the full research process. 3) Distilling these trajectories into a smaller Nemotron-3-Nano-30B-A3B model via long-context SFT on 8 H100 GPUs. The resulting OpenResearcher-30B-A3B model achieves state-of-the-art performance on benchmarks like BrowseComp-Plus, demonstrating that synthetic research experience generalizes to the live web. The post highlights a shift in compute towards synthetic experience generation and notes adoption by NVIDIA.
  3. Your coding harness shouldn't be a black boxThis post introduces NCode, an open-source AI model harness developed by Noumena. It highlights the 'black-box' problem of proprietary harnesses and the benefits of an inspectable and modifiable infrastructure layer. NCode allows users to bring their own models (open-weight or proprietary), fine-tune them, and integrate them with tools for code editing and execution. The post emphasizes the importance of vertically integrating the model with the harness for productivity and reliability, showcasing how NCode was used internally for six months on millions of lines of code and dozens of services. It also details NCode's flexibility in model integration (e.g., OpenAI API, OpenAI-compatible endpoints, self-hosted models) and its tiered intelligence structure based on real-world usage. The post concludes by linking NCode integration with Lambda's GPU serving capabilities for open-weight models.
  4. GLM 5.2: a new rise of open-weight agentic modelsThis post highlights GLM 5.2 as a significant advancement in open-weight agentic models, demonstrating performance comparable to proprietary models on various benchmarks. It details how GLM 5.2 is being adopted as a capable subagent in complex AI pipelines, such as autoresearch tools, and discusses the trade-offs between running large models on-premises versus cloud infrastructure, emphasizing that serving these models at full quality requires substantial infrastructure investment. The post positions GLM 5.2 as a tipping point for the usability of open-source models in everyday engineering workflows.