Recipe-based model launches
Uses model-specific recipes containing model IDs, GPU requirements, setup steps, and serving flags; `iwant up MODEL` selects the latest recipe version.
iwant
iwant is an open-source CLI for launching open-weight language models on bare-metal GPUs in your Google Cloud account, with an OpenAI-compatible API endpoint.
What is iwant?
iwant is an open-source command-line tool for launching open-weight language models on bare-metal GPUs in a user's own Google Cloud account. It uses model recipes to select hardware and serving configuration, provisions the requested capacity, downloads and loads the model, and waits for the OpenAI-compatible models API to respond.
After a launch, iwant provides a server URL and API key that can be used with curl or an OpenAI-compatible application. Users can preview a plan, choose capacity options, inspect running deployments, connect over SSH, and remove clusters when finished. Cloud infrastructure is billed to the user's account until teardown, so the tool is intended for users who can manage Google Cloud credentials, quota, capacity, and resource lifecycle.
Key features
Cloud GPU provisioning
Provisions bare-metal GPU capacity in the user's Google Cloud account and configures the serving environment without requiring a separate hosted inference account.
OpenAI-compatible endpoint
Checks that the models API responds before reporting the server ready, then prints a server URL and API key for client configuration.
Launch controls
Supports `--dry-run` for plan inspection, `--spot` for spot capacity, region selection, custom idle timers, quiet output, and optional API keys.
Deployment management
Provides `iwant list`, `iwant ssh`, `iwant auth`, and `iwant down` for checking deployments, connecting to running clusters, validating credentials, and tearing resources down.
Model download and discovery options
Can pass `HF_TOKEN` for authenticated or gated model downloads, and supports finding the served model ID through the `/models` endpoint.
Use cases
Testing an open-weight model
Preview the recipe and capacity plan, provision a GPU in GCP, and send prompts through the returned URL and key while keeping infrastructure in the user's cloud account.
Connecting an application to self-managed inference
Point an existing OpenAI-compatible application at the generated base URL and use the served model ID without changing the application's general request pattern.
Running larger model configurations
Use recipes with larger multi-GPU configurations, such as eight H100 or H200 GPUs, when the selected model's documented requirements exceed a single-GPU setup.
Short-lived GPU experiments
Use `--spot`, region selection, dry runs, and configurable idle teardown when experimenting with capacity choices and managing the lifetime of temporary deployments.
Managing active clusters
Inspect deployments with `iwant list`, connect using `iwant ssh`, and remove named clusters with `iwant down` when operational access or cleanup is required.
Pros and Cons
Pros
- Launches open-weight models through a concise CLI workflow.
- Returns an OpenAI-compatible endpoint suitable for existing clients and curl requests.
- Recipe files capture model-specific hardware and serving requirements.
- Includes dry runs, spot capacity, idle teardown settings, SSH access, and cluster management commands.
Cons
- Current documentation identifies GCP as the supported cloud, so other providers are not established by the available evidence.
- GPU provisioning, model downloads, and loading take time; spot capacity can be reclaimed.
- Users are responsible for cloud billing and explicitly tearing down resources when finished.
FAQ
What is required to launch a model?
Install iwant with the provided shell installer, authenticate with the Google Cloud CLI, and use a Google Cloud project with billing enabled, L4 quota, and available capacity. Python 3.10+ is supported when installing from source.
How does a model launch work?
Run `iwant up MODEL` to provision a GPU, configure the server, and wait for the models API to respond. The command prints a server URL and API key for use with an OpenAI-compatible client or direct API requests.
Which cloud provider does iwant support?
The current supported cloud is GCP. You can select a region, request spot capacity, and use the launch plan to preview resource requirements before provisioning.
How do I stop a deployment and control costs?
Run `iwant down` to choose a cluster interactively, or run `iwant down EXACT_CLUSTER_NAME` for scripts. Cloud resources continue billing in the user's account until they are removed.
Which models and GPU configurations are available?
The repository includes recipes for models including gpt-oss-20b, DeepSeek variants, mimo-v2.6-flash, ling-3.0-flash-fp8, and step-3.7-flash variants. Recorded hardware ranges from one L4 GPU to eight H100 or H200 GPUs, but repository evidence is not a guarantee for every launch.
Quick Facts
- Category
- Developer Tool
- Interface
- Command-line interface
- Supported cloud
- Google Cloud Platform (GCP)
- Installation
- Shell installer or Python 3.10+ from source
- Endpoint format
- OpenAI-compatible API
- Source
- github.com/narevai/iwant
iwant Alternatives
AakarDev AI
AakarDev AI helps teams manage AI provider access, project setup, logs, and analytics in one dashboard. BYOK support included.
ByteAsk
ByteAsk is a terminal-first AI coding agent for C and C++ that edits repos and verifies changes with compilers, debuggers, sanitizers, and tests.
Epherem
Epherem scans web apps built with Lovable, Bolt, v0, Replit, Cursor, and Claude Code for security flaws in source code and dependencies, without running the app.
CreateOS Sandbox
CreateOS Sandbox is an isolated compute environment for running code and agent workloads in Firecracker micro-VMs with private networking and SDK, CLI, or MCP control.
GitBot
GitBot is a local browser workspace for reusable coding bots powered by Claude Code, Codex, or OpenCode, with persistent threads and configurable permissions.
hob
hob is an independent workspace for coding agents, with local control over sessions, terminals, history, routing, and follow-up work.