Written by Jakub Rusinowski · Last updated September 19, 2026
Model library → LFM2.5 → LFM2.5 2.6B
Liquid AI's on-device agentic model: 2.69B parameters over 30 layers, of which only 8 are attention — the other 22 are double-gated short convolutions that cache nothing. That is why its KV cache stays small at a 128K context where a conventional 2.7B model's would not. Trained on 34T tokens, post-trained for tool calling.
LFM2.5 2.6B needs about 2 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
| Parameters | 2.69 Billion |
| Context window | 131,072 |
| Architecture | Hybrid (22 conv + 8 GQA) |
| Provider | Liquid AI |
| Licence | LFM Open License v1.0 |
| Specified at | Q4_K_M |
| System RAM | 8 GB |
| Record updated | 2026-09-19 |
Corroborated — Two or more independent sources agree on these figures, but the model card itself was not retrieved. Treat the numbers as good rather than confirmed.
LFM Open License v1.0 — commercial use permitted. Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Modelled on a reference NVIDIA RTX 4090 (24 GB). Weights plus framework overhead only — this model publishes no architecture we can read, so no KV cache is included. A real session needs more; the figure is a floor, not a target. Speed figures are ESTIMATES from the memory-bandwidth roofline described on the methodology page, not benchmarks we ran — rows marked measured come from published or reader-submitted runs. VRAM here includes the KV cache, so it reads higher than the headline figure above, which does not.
| Quant | Bits/weight | Weights | VRAM needed | Est. speed | Fit on 24 GB |
|---|---|---|---|---|---|
| Q2_K | 2.63 | 0.9 GB | 1.7 GB | ~253 tok/s (est.) | Fits comfortably |
| Q3_K_M | 3.41 | 1.1 GB | 1.9 GB | ~233 tok/s (est.) | Fits comfortably |
| Q4_K_M | 4.83 | 1.6 GB | 2.4 GB | ~203 tok/s (est.) | Fits comfortably |
| Q5_K_M | 5.67 | 1.9 GB | 2.7 GB | ~189 tok/s (est.) | Fits comfortably |
| Q6_K | 6.56 | 2.2 GB | 3 GB | ~176 tok/s (est.) | Fits comfortably |
| Q8_0 | 8.50 | 2.9 GB | 3.7 GB | ~152 tok/s (est.) | Fits comfortably |
| F16 | 16.00 | 5.4 GB | 6.2 GB | ~101 tok/s (est.) | Fits comfortably |
Want to set your own context length and KV-cache quantization? Use the interactive VRAM calculator.
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
The cheapest catalogued GPU that runs LFM2.5 2.6B is the Intel Arc B570 (10 GB).
Install Ollama, then run:
ollama run lfm2-5Weights on Hugging Face: LiquidAI/LFM2.5-2.6B.
Best for: on device, agents, tool calling, edge.
← All LFM2.5 models | VRAM calculator | Check your own hardware