badtheorylabs/Macaw
Text Generation • 3B • Updated • 1.24k • 9
AI & ML interests
None defined yet.
Recent Activity
Organization Card
BTL
Independent research lab in Lagos.
Intelligence efficient enough to own.
We build frontier models efficient enough to run on hardware people already own.
Models
| Model | What it is |
|---|---|
| BTL-4 | 35.1B mixture-of-experts, ~2.1B active parameters per token, 262K context |
| BTL-4 Compact | The same model in one 9.96 GB GGUF at 2.30 bits per weight, retaining 94.1% of measured full-precision behaviour (111 of 118 on a teacher-correct replay set). Runs in stock llama.cpp, Ollama, and LM Studio. |
| BTL-3 | 27B agentic coding and tool-use model |
| BTL-3 Compact | The complete text model in one 8.39 GB native GGUF |
| Macaw | 2.7B on-device macOS agent, 97 verified tools, no cloud |
Elsewhere
- Research index: badtheorylabs.com/papers
- Thesis: badtheorylabs.com/thesis
- Code: github.com/Badtheorylabs
models 7
badtheorylabs/BTL-4-Compact
Text Generation • 35B • Updated • 79.9k • 62
badtheorylabs/BTL-4
Text Generation • 35B • Updated • 9.45k • 85
badtheorylabs/Macaw-4bit-MLX
Text Generation • 0.4B • Updated • 719 • 15
badtheorylabs/BTL-3-Compact
Text Generation • 8B • Updated • 376 • 42
badtheorylabs/BTL-3
Text Generation • Updated • 45 • 73
badtheorylabs/btl-2-coder
Text Generation • Updated • 5 • 4
datasets 0
None public yet