hipfire-LFM2.5-1.2B

License: LFM Open License v1.0 — the LICENSE file in this repo is copied verbatim from LiquidAI/LFM2.5-1.2B-Instruct. This is a quantized derivative of a model created and released by Liquid AI; all upstream attribution and license terms apply.

hipfire-native Q8 quantization of LiquidAI/LFM2.5-1.2B-Instruct — a 1.2B-parameter hybrid edge model: 16 layers (10 gated short-conv + 6 GQA full-attention, 32 q heads / 8 kv heads), hidden 2048, SwiGLU FFN 12288, vocab 65536, 128K context.

The .q8 file runs with the hipfire inference engine (HIP/ROCm-direct, Rust, no Python in the hot path). It is not GGUF/safetensors and won't load in llama.cpp / transformers. All weights are stored at Q8 and dequantized to f32 on the fly inside the GEMV kernels.

Files

File	Quant	Size
lfm2.5-1.2b-instruct.q8	Q8 (all tensors)	1.25 GB

The LiquidAI chat template is embedded in the file and used by default (jinja chat is default-on in hipfire; opt out with HIPFIRE_JINJA_CHAT=0).

Usage (hipfire)

# hipfire CLI (registry tag, once available in your hipfire version):
hipfire pull lfm2.5:1.2b
hipfire run lfm2.5:1.2b "What is the capital of France?"

# or download directly and point at the file:
hf download hipfire-models/hipfire-LFM2.5-1.2B lfm2.5-1.2b-instruct.q8 --local-dir ~/.hipfire/models
hipfire serve --model ~/.hipfire/models/lfm2.5-1.2b-instruct.q8

Validation

Greedy (temperature 0.0) daemon transcript on RDNA3 (gfx1100 / RX 7900 XTX), hipfire v0.2.1:

prompt: "What is the capital of France?"
output: "The capital of France is **Paris**."
409 tok/s decode, 45 ms prefill

arch_id 11 in the hipfire HFQ header.