Two open models, both local, both private, both Apache 2.0. Grab the single-file Terse Micro model, or the Orchid 1.0 weights.
Available now
A ~423M ternary-weight model in a single ~182 MB file. Trained from scratch, runs on any CPU — no GPU, no cloud.
orchid
Model · fine-tune
Our 2B model — a fine-tune of Microsoft BitNet b1.58, ORPO-aligned. Weights and code are public; run it locally with the ternative engine.
⚑ Terse Micro is hosted on Hugging Face; Orchid 1.0's weights and code live on GitHub. Both run fully offline.
No account, no cloud, no data leaving your device. Works offline.
Both models run on hardware you already own — Terse needs only a CPU.
Free for personal and commercial use. Weights and engine are public.
Grab the raw weights and run them yourself. Terse Micro is a single-file download; Orchid needs its base plus the LoRA adapter and a ternative build.
Model · weights
The whole model in one ~182 MB file (TQ2_0 GGUF). Runs on a CPU via the terse-arch llama.cpp fork.
# single file, from Hugging Face huggingface-cli download MicheRomChis/micro-terse \ terse-micro.TQ2_0.gguf --local-dir ./terse # then run with the terse-arch llama.cpp fork ./llama-cli -m ./terse/terse-micro.TQ2_0.gguf
Model · weights
The I2_S base GGUF (~1.1 GB) plus the aligned LoRA adapter (~90 MB), hosted on Hugging Face.
huggingface-cli download MicheRomChis/orchid-1.0 \
ggml-model-i2_s.gguf dpo_aligned-lora.gguf \
--local-dir ./orchid-models
Clone and build the C++17 / CUDA engine, then point it at the Orchid weights. Windows & Linux.
git clone --depth 1 \ https://github.com/michelangeloromerochisco/ternative cmake -B build -DCMAKE_BUILD_TYPE=Release \ && cmake --build build --parallel
Both models were built to run on consumer hardware. No discrete GPU is required — Terse Micro runs on CPU alone, and Orchid runs in ~4 GB VRAM or CPU-only.
| Minimum | Recommended | |
|---|---|---|
| GPU VRAM | 0 (CPU) | 4 GB |
| RAM | 8 GB | 16 GB |
| Storage | 1.3 GB | 2 GB |
| OS | Win / Linux | — |