[How-To] Music Models on 4GB, Serverless Agents on Bedrock, and Self-Building AI

March 5 Office Hours Recap: Open-weight music generation hits consumer hardware, OpenClaw goes serverless on AWS, and a personal agent turns meetings into deployed software.

Stanislav Huseletov, Stephen Barr, and Leonardo Gonzalez

Mar 05, 2026

This session covered three distinct technical frontiers: running open-source music generation on hardware you already own, deploying OpenClaw as a serverless agent on AWS Bedrock without standing up infrastructure, and using a personal agent to autonomously convert meeting discussions into prototypes and content.

1. Open-Source Music Generation on Consumer Hardware

Leonardo Gonzalez demoed an open-weight music model built by Ace Studio that generates full songs from text prompts on modest hardware.

Technical specs:

Runs on as little as 4GB of VRAM (consumer GPUs, Apple Silicon)
Completely open-weight and open-source
Ships with a Gradio web app for text-to-music generation
Supports LoRA adapters for fine-tuning rhythms, vocal style, and sound
Built on Qwen 2.5 and Gemini 2.5 foundations

The standout feature: LRC file generation. When the model generates a song, it also outputs timestamped lyrics in LRC format. These can be exported directly as subtitles for CapCut, YouTube, or any video editor. No manual transcription. No alignment tools. It ships out of the box.

Leonardo converted the model’s technical report to PDF, fed it to ChatGPT, and had it compose a song about the model itself. The output had coherent lyrics aligned to the beat with exportable captions.

[Deep Dive] Qwen 3.5 Brings Native Multimodality and Long Context to Small Open Models

Leonardo Gonzalez

Mar 4

Read full story

Why it matters: This model is built on Qwen 2.5. If the next version uses the Qwen 3.5 (with its native multimodality and 262K context), the quality jump would be significant. Open-weight licensing means anyone can fine-tune it for their voice, genre, or use case.

2. Serverless OpenClaw on AWS Bedrock via Agent Core

Stephen Barr deployed a fully functional OpenClaw agent on AWS using one prompt. No standing infrastructure. No manual configuration.

The prompt: “Read the README and help me get it set up.”

What the agent did from that single instruction:

Built a deployment to-do list from the repo documentation
Configured AWS credentials
Created a Telegram bot and stored the token in AWS Secrets Manager
Set up the webhook
Deployed 8 CloudFormation CDK stacks for routing and infrastructure

The architecture:

Agent Core (built on EKS) launches ephemeral Docker containers when you message the agent and shuts them down when idle
S3 stores all agent files, memories, and workspace data
AWS Secrets Manager handles API key storage (addressing the common criticism that OpenClaw writes keys in plain text)
Connected to Opus on Bedrock by default, swappable to Kimi K2.5 or any Bedrock-supported model

The AWS-native advantage: Because this OpenClaw lives inside AWS, operations like downloading videos with yt-dlp go directly to S3 at internal network speed. No local download, no upload round trip.

Stan asked if the agent could provision EC2 instances. Stephen confirmed: modify the IAM policy, create an OpenClaw skill, and the agent handles provisioning on its own.

Stephen uses Whisper Flow for voice input, dictating instructions instead of typing.

Bedrock Agent Core Repository
Setup prompt: “Read the README and help me get it set up”

3. Cooperate: Internal Enterprise OpenClaw Deployment

David Proctor built Clawporate, an internal tool for one-click OpenClaw deployment within the company. Trilogy or DevFactory domain holders can log in, spin up an isolated gateway, and optionally upload an existing OpenClaw configuration.

The security model that emerged from real usage:

David runs two OpenClaw instances. His personal one (local, broad access) and his work one (deployed via Cooperate, limited scope). The personal agent can securely update the remote one. The remote one cannot access the personal one’s data.

He demonstrated this from his phone: messaged his local OpenClaw with “transfer this credential securely to the remote one.” The local agent stored the credential in AWS Secrets Manager and updated the remote agent’s configuration file. Controlled, one-directional exposure.

David also pointed out that you can tell your OpenClaw to read Rahul Subramaniam’s recently published article on OpenClaw real-world practices and implement the recommendations. I confirmed he did exactly that: asked his agent to evaluate the article first, then implement it.

4. The Shadow Agent: Meetings to Deployed Software

I described Shadow, a personal OpenClaw agent with read-only access to call transcripts, meeting notes, articles, and personal documents.

Shadow’s workflow:

Ingests new meeting content automatically
Classifies content into ideas, buildable projects, or publishable articles
Builds and deploys prototypes to Vercel without asking
Drafts articles and emails them for review

I asked the group: “If you’re not using OpenClaw, what would make you use it? Would you use Shadow if you had one?”

5. Quick Technical Hits

Prompt repetition: I shared a finding from Google researchers showing that repeating the same prompt verbatim to non-reasoning models improves output quality. He confirmed this works for reasoning models too. If the first output is borderline, repeat the exact same prompt.

Published this week: