Meet Tamebi: A General-Purpose Python Package for Building with Open-Source AI
There's a gap between wanting to use open-source AI and actually using it. Not a knowledge gap. Not an interest gap. An infrastructure gap. Developers spend hours figuring out which models fit their hardware, how to wire up inference, how to build agents that actually work in production. The tooling for open-source AI has been fragmented, hardware-specific, and hard to reason about. This package is our answer to that.
What is Tamebi's package ?
Tamebi is a general-purpose Python package and CLI for working with open-source AI models. From understanding what you can run, to running it, to building on top of it. Think of it the way you think about the openai package, but for the open-source ecosystem: a single, coherent library that grows with you from your first experiment to production deployment.
We're building it feature by feature, in public, and we're just getting started.
The first feature: know what you can run
One of the the hardest parts of getting started with local AI isn't motivation. It's uncertainty.
Can my machine run a 7B model? What about 14B? Should I use FP16 or INT4? What happens to memory when I extend context to 32K? Can I serve multiple users at once?
These questions have real answers, but finding them today means cross-referencing benchmarks, reading GitHub issues, and doing mental math. That's friction that stops people before they start.
So we built tamebi check: one command that automatically scans your hardware and tells you exactly which models fit, at which precision, with what performance to expect.
pip install tamebi
# or, faster:
uv pip install tamebi
tamebi checkIn seconds you get :
A full hardware profile (CPU, RAM, GPU/VRAM, OS)
A ranked compatibility matrix across INT4, INT8, and FP16
Estimated tokens/sec and time to first token
Ready-to-copy Ollama commands for every compatible model
Multi-user serving estimates with
--batch-sizeJSON output for automation and scripting
It works automatically on NVIDIA, AMD, Apple Silicon, and CPU-only systems, with no config, no flags, no guessing.
This is v1. It already saves hours of trial and error. But it's the foundation, not the ceiling.
What's coming ?
Hardware detection is just the entry point. We're building Tamebi into something much broader:
Model inference: run open-source models directly through the package, with a clean, consistent API
Agent primitives: tools, memory, and orchestration for building AI agents on top of local or remote models
Provider abstraction: switch between local inference, cloud APIs, and hybrid setups without rewriting your code
Evaluation and benchmarking: understand model quality, not just model fit
Deployment helpers: move from local prototype to edge or server deployment with less friction
The goal is for Tamebi to be the package you reach for whenever you're building with open-source AI, whether you're a student running your first model, a developer shipping a product, or a team planning infrastructure.
Who is it for ?
Developers building AI-powered apps
Researchers evaluating models before committing to one
Students exploring local inference for the first time
Startup teams sizing their stack before they scale
Platform engineers planning edge deployments
Anyone who wants to work with open-source AI without fighting the tooling
Docs at lab.tamebi.ai/docs. The model catalog covers models from Meta, Mistral, Google, Qwen, DeepSeek, Kimi, Liquid, AllenAI, and more, and updates weekly.
This is the beginning. We're building in public, and we'd love to hear what you want Tamebi to do next.