---
license: apache-2.0
base_model: SpectraSuite/TriLM_3.9B_Unpacked
tags:
- ternary
- tq2_0
- gguf
- ppc64le
- ibm-power
- llama.cpp
---

# TriLM 3.9B — TQ2_0 GGUF for IBM Power

[TriLM 3.9B](https://huggingface.co/SpectraSuite/TriLM_3.9B_Unpacked) (Spectra Suite,
ternary {-1, 0, +1} weights) quantized to **TQ2_0** (2.06 bits per weight) for fast
CPU inference with llama.cpp — optimized for IBM POWER9 and later with the
[LibrePower VSX ternary kernels](https://linux.librepower.org).

- **File**: `TriLM-3.9B-TQ2_0.gguf` (1.46 GiB, 3.99 B parameters)
- **Quantization**: TQ2_0 — exact ternary, no quality loss vs the original ternary weights
- No GPU required.

## Run it on IBM Power (ppc64le)

```bash
curl -fsSL https://linux.librepower.org/install.sh | sudo sh
sudo apt install librepower-llama

lp-llama-completion -m TriLM-3.9B-TQ2_0.gguf -p "Once upon a time" -n 64 -t 48
```

Works with any recent llama.cpp on any architecture; the LibrePower build adds
VSX acceleration on Power (4.3x prompt / 2.4x generation vs the generic path).

## AIX / big-endian

`TriLM-3.9B-TQ2_0-be.gguf` is the **big-endian** variant for IBM AIX
(`dnf install llama-aix` from [aix.librepower.org](https://aix.librepower.org)):

```sh
lp-llama-completion -m TriLM-3.9B-TQ2_0-be.gguf -p "Once upon a time" -n 64 -t 16
```

## Performance (IBM POWER9, Ubuntu 22.04, llama-bench)

| Test | Threads | Tokens/s |
|------|---------|----------|
| Prompt processing (pp64) | 96 | 121.2 |
| Generation (tg64) | 48 | 29.5 |
| Generation (tg64) | 24 | 22.3 |

Note: this is a **base model** (no instruction tuning) — use completion-style
prompts, not chat.

## Credits

- Model: [Spectra Suite](https://huggingface.co/SpectraSuite) (Apache 2.0)
- TQ2_0 quantization & Power packaging: [LibrePower](https://librepower.org)