hipfire-native Q8 quantization of
LiquidAI/LFM2.5-1.2B-Thinking —
the reasoning variant of LFM2.5-1.2B. Same hybrid edge architecture: 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. Emits <think>…</think>
reasoning before the final answer.
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-thinking.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).
The template strips past-turn <think> blocks from history by default
(keep_past_thinking=false), matching upstream behavior.
Usage (hipfire)
# hipfire CLI (registry tag, once available in your hipfire version):
hipfire pull lfm2.5:1.2b-thinking
hipfire run lfm2.5:1.2b-thinking "What is the capital of France?"
# or download directly and point at the file:
hf download hipfire-models/hipfire-LFM2.5-1.2B-Thinking lfm2.5-1.2b-thinking.q8 --local-dir ~/.hipfire/models
hipfire serve --model ~/.hipfire/models/lfm2.5-1.2b-thinking.q8
Give it headroom: reasoning turns typically need a few hundred tokens of
max_tokens before the final answer appears after </think>.
prompt: "What is the capital of France?"
output: "<think> Okay, let's see. The question is asking "What is the capital
of France?" … The capital of France is indeed Paris. The Eiffel Tower
is there, the Louvre Museum, Notre-Dame Cathedral… </think>
The capital of France is **Paris**.
Paris serves as the country's political, cultural, and economic
center, renowned for landmarks like the Eiffel Tower, Louvre Museum,
and Notre-Dame Cathedral.
**Answer:** The capital of France is Paris."
~395 tok/s decode, 45 ms prefill
Known cosmetic nit: at the very end of a greedy completion the model may emit a
short run of byte-fallback tokens (an emoji) that stream as � replacement
characters in clients that decode token-by-token; the answer text itself is
unaffected.
