Punk Band of Software Development - Vibe Coding with All Hands AI, Devstral, and vLLM
- Chandan Kumar
- Jun 30
- 6 min read
Updated: Sep 27
On July 16, 2025, Chandan Kumar hosted an in-depth webinar exploring the rising concept of “Vibe Coding”—a term popularized by Andrej Karpathy and increasingly debated across developer circles. The session went far beyond definitions, diving into philosophy, industry debates, practical use cases, and even a live demo of open-source tools that can rival paid AI coding assistants.
If you missed it, here’s the complete recap.
What Is Vibe Coding?
The term Vibe Coding originated from a tweet by Andrej Karpathy. At its core, it describes a new way of programming powered by large language models (LLMs), where developers (and even non-developers) can “go with the flow.”
Instead of obsessing over clean architecture, performance optimization, or purist coding practices, vibe coding focuses on outcomes, not methodology. You prompt an AI assistant like ChatGPT, Cursor, Sonnet, or GitHub Copilot, and let it generate code for you. You copy, paste, iterate, and vibe your way toward a working solution.
“It’s like punk rock for software development. You don’t have to be a classically trained musician to play punk rock—you just plug in your guitar and play. Vibe coding does the same for programming.” – Chandan Kumar
This framing has sparked both excitement and criticism. Supporters see it as democratizing software creation; detractors dismiss it as sloppy “bug coding.”
Punk Rock Meets Programming
Chandan drew a compelling analogy between vibe coding and the history of music.
Traditionally, becoming a developer required formal training—he shared his own journey of months-long embedded C programming courses before he could contribute meaningfully as an engineer. Similarly, classical musicians once had to train rigorously to be taken seriously.
Punk rock changed that. Artists without formal music education picked up instruments and created music that resonated, even if it broke traditional rules.
Vibe coding is that same cultural moment for software. Non-developers, product managers, or technologists without deep programming skills can now build working applications with the help of AI tools. They may not write “clean code,” but they can ship prototypes, apps, or agents that solve real problems.
The Debate: Fad or Future?
Like punk rock, vibe coding has its critics and defenders.
The Criticism
“Not real code” – Purists argue that AI-generated code is messy, unoptimized, and full of bugs.
Reliability concerns – LLMs are probabilistic, not deterministic. They can hallucinate APIs or produce insecure code.
Not production-ready – Relying solely on AI is risky for mission-critical software.
The Defense
Rapid prototyping – Perfect for spinning up agents, scrapers, or scaffolding in minutes.
Language accessibility – Developers can work in unfamiliar languages (e.g., Scala, Go) by prompting the AI.
Empowerment – Non-developers in technical teams can build small tools or automations without full coding expertise.
Augmentation, not replacement – AI doesn’t eliminate developer jobs; it makes them more productive.
“Vibe coding doesn’t replace developers—it augments them. It’s autocomplete on steroids, not the end of programming.” – Chandan Kumar
Tools of the Trade
Chandan walked through the main players shaping the vibe coding ecosystem:
VS Code – Still the base editor for many, with plugins like Copilot.
Cursor – A powerful AI-first coding IDE, responsive and production-ready.
Kiro (AWS’s play) – Recently launched, integrates flow and vibe-based coding.
Sonnet, Windsurf – Alternative IDEs with different agents and approaches.
OpenHands (by All Hands AI) – An open-source, self-hosted alternative.
A recurring theme: most commercial tools are paid, with token limits that quickly push users into $20–200/month subscriptions. This creates space for open-source, self-hosted options.
What You’ll Discover in This Post
Understanding Vibe Coding – What it means and why it's the next frontier in software development.
Overview of All Hands AI – Key features, security-first design, and integration benefits with Devstral and VLLM.
A detailed setup guide – From GPU provisioning to model serving with step-by-step instructions.
Visual documentation – Screenshots, dashboards, and UI views to aid setup and monitoring.
Real-world scenarios – Practical examples and productivity tips to harness the full power of this AI toolchain.
Vibe Coding: Rethinking the Developer Experience
Traditional coding often means hours of isolated work, repetitive debugging, and generating boilerplate code. But with AI, there's a chance to shift from monotony to flow. Vibe Coding centers around harnessing intelligent tools to build an experience that is seamless, creative, and deeply satisfying.
Imagine coding alongside an AI assistant that:
Proposes contextual optimizations proactively.
Detects subtle bugs before they become blockers.
Automatically writes boilerplate, tests, and structured refactors.
Helps teams stay in sync by suggesting improvements across shared codebases.
Vibe Coding is about staying in a state of flow—where deep work becomes natural. Tools like All Hands AI serve as intelligent, context-aware copilots that understand your intent and architecture, helping teams focus on what truly matters: building exceptional software.
Meet All Hands AI: Secure, Customizable, and Developer-Centric
In an age dominated by cloud-first services, All Hands AI is a powerful counterpoint. It’s a fully self-hostable, open-source AI coding assistant built with privacy, control, and extensibility in mind—ideal for teams wary of proprietary tools like GitHub Copilot or Cody.
What Sets It Apart?
Full Self-Hosting: Keep your codebase and LLM processing within your own infrastructure. Essential for compliance-heavy industries.
Tailored Integrations: Configure it to your team’s specific workflows and even fine-tune it with private repo embeddings.
Data Sovereignty: Avoid vendor lock-in. Your intellectual property stays secure and within your control.
Think of it as a company-specific Copilot—trained on your internal best practices, libraries, and unique project DNA.
The Power Trio: All Hands AI + Devstral + VLLM on Denvr Cloud
Pairing All Hands AI with the right backend tech stack transforms it from helpful to indispensable. Enter Devstral, a purpose-built LLM for software engineering, and VLLM, a high-performance inference engine. Hosted on Denvr Cloud, this trio creates a best-in-class AI development platform.
Devstral: LLM Purpose-Built for Engineers
Created in partnership with Mistral AI, Devstral is optimized for agentic workflows and coding complexity:
Navigates codebases intelligently, identifying dependencies and relevant snippets.
Edits across multiple files, enabling holistic code refactors.
Drives AI agents that handle automation, testing, and static analysis.
Its strong performance on the SWE-bench benchmark cements it as one of the most capable open LLMs for dev teams.
VLLM: Accelerated LLM Inference
Built for scale and speed, VLLM ensures fast responses from large models like Devstral. Thanks to techniques like PagedAttention, it offers:
High throughput
Low latency
Efficient GPU memory usage
It’s the perfect engine to fuel real-time interactions with your AI assistant.
Denvr Cloud: Optimized Hosting for AI Workloads
Want to deploy your stack at scale? Denvr Cloud offers:
🚀 Top-tier GPUs like A100, H100, and Gaudi2.
🛠️ Container-native workflows for CI/CD, Docker, and Kubernetes.
🔐 Enterprise-ready isolation, perfect for regulated teams.
Whether you're prototyping solo or scaling across an organization, Denvr Cloud is a battle-tested choice.
Deployment Prerequisites
Before diving in, make sure you have:
A Denvr Cloud account with GPU resources.
(Optional) A Hugging Face account for model access.
Step-by-Step Guide to Deploying the Stack
1. Provision GPU VM on Denvr Cloud
Choose an A100 or H100 instance.
Recommended OS: Ubuntu 22.04.
Ensure Docker and SSH access are enabled.
Add the user to the Docker group:
sudo usermod -a -G docker ubuntu
Clone the deployment repo:
git clone https://github.com/denvrdata/examples.git
2. Install VLLM and Dependencies
pip install --upgrade vllm pyopenssl
3. Launch Devstral with VLLM
screen -S vllm
vllm serve mistralai/Devstral-Small-2505 \
--tokenizer_mode mistral \
--config_format mistral \
--load_format mistral \
--tool-call-parser mistral \
--enable-auto-tool-choice \
--tensor-parallel-size 8
Check availability with:
python3 llm_ping.py <IP>
Monitoring with Prometheus and Grafana
From the examples/all-hands-ai-mistral/ folder:
docker compose up -d
Access Grafana at http://<IP>:3000
Add Prometheus as a data source
Import dashboards:
vllm_dashboard.json
NVIDIA DCGM (GPU metrics)
Running the OpenHands App
Launch the app with:
docker run -it --rm --pull=always \
-e SANDBOX_RUNTIME_CONTAINER_IMAGE=docker.all-hands.dev/all-hands-ai/runtime:0.38-nikolaik \
-e LOG_ALL_EVENTS=true \
-v /var/run/docker.sock:/var/run/docker.sock \
-v ~/.openhands-state:/.openhands-state \
-p 3001:3000 \
--add-host host.docker.internal:host-gateway \
--name openhands-app \
docker.all-hands.dev/all-hands-ai/openhands:0.38
Note: Grafana uses port 3000, so OpenHands is mapped to 3001.
Visit: http://<IP>:3001
Configure Custom LLM Inference
Within the app’s settings:
Model: openai/mistralai/Devstral-Small-2505
Base URL: http://<your-server>:8000/v1
API Key: token
Example: All Hands AI in Action
Prompt
Build a To-Do app using FastAPI and React, with: Task add/delete/complete features One-page layout SQLite backend
Output
A fully scaffolded, cleanly written codebase generated instantly by your AI pair programmer.
Real-World Applications
Mass syntax corrections in monorepos
Custom security rule audits
Test generation for legacy code
Repository-aware smart suggestions
Conclusion
When you combine All Hands AI with Devstral and VLLM—especially on Denvr Cloud—you’re creating more than a Copilot alternative. You’re building a secure, extensible, AI-powered development platform tailored to your team’s unique needs.
Comments