Measured model
Measured on a Mac Studio (M5 Max, 36GB)LFM2.5-8B-A1B
Liquid AI · Ollama name hf.co/LiquidAI/LFM2.5-8B-A1B-GGUF:Q4_K_M
On a Mac Studio (M5 Max, 36GB), LFM2.5-8B-A1B answers a simple question in about 1.7 s, gets 71% of the Japanese knowledge questions right, 78% of the math problems, and uses 5.4GB of memory.
Results
| Measure | Result | Among measured models | Note |
|---|---|---|---|
| Reply time | about 1.7 s | 8 of 33 | Until one simple common-sense question is answered (median of 50) |
| Knowledge (high school to university) | 71% | 28 of 33 | 200 questions, ±7, 3 cut off (counted wrong) |
| Math (word problems) | 78% | 28 of 33 | 100 questions, ±9 |
| Memory used | 5.4GB | 8 of 33 | |
| EN→JA translation | 29 pts | 24 of 30 | out of 100 (chrF), about 1.5 s a sentence |
| JA→EN translation | 51 pts | 25 of 30 | out of 100 (chrF), about 1.5 s a sentence |
"Among measured models" is the place among the 33 models measured on this site (lower is better for time and memory).
Try it on your Mac
- Install the Mac version of Ollama from ollama.com and start it.
- Type this in Terminal. The first time, the model downloads; then you can chat with it.
ollama run hf.co/LiquidAI/LFM2.5-8B-A1B-GGUF:Q4_K_M
It used 5.4GB here. To use it comfortably alongside other apps, a Mac with 16GB of memory or more is a good guide (the model then takes about half the memory).
About the model
- Maker
- Liquid AI
- Size
- 8.47B parameters
- Quantization
- Q4_K_M
- Thinks before answering
- Yes
- Longest input at once
- 128,000 tokens
- Released (Hugging Face)
- May 28, 2026
- Generation speed
- 237.2 tok/s
- Start-up (loading)
- about 0.6 s
- Measured on
- Oct 10, 2026
- Ollama
- 0.34.3
- Hugging Face
- LiquidAI/LFM2.5-8B-A1B
- Series
- LFM
What the parts of a name mean (size, quantization and so on): how to read a model name.
Models of similar memory
- gemma-4-E4B-it (5.4GB of memory, knowledge 79%, reply about 2.6 s)
- Ornith-1.0-9B (6.2GB of memory, knowledge 85%, reply about 3.8 s)
- Ornith-1.5-9B (6.4GB of memory, knowledge 77%, reply about 2.5 s)
- DeepSeek-R1-0528-Qwen3-8B (6.4GB of memory, knowledge 72%, reply about 9.4 s)
- NVIDIA-Nemotron-3.5-Bolt-E4B-BF16 (6.5GB of memory, knowledge 76%, reply about 2.6 s)
Compare the main models on the text page. The series page lists every model measured.
Task recommendations and capabilities · Measurements by model