LFM2.5-8B-A1B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
8.3B total parameters resident — about 5.8 GB at Q4_K_M, comfortable on a 6 GB card or a phone with enough RAM — with 1.5B active per token. Note the name understates it: "A1B" is a naming convention, the measured active count is 1.5B. 24 layers, 18 double-gated LIV convolution blocks plus 6 GQA attention layers, which is why its KV cache is far smaller than a 24-layer transformer's. 128K context, vocabulary widened from 65,536 to 128,000 for non-Latin scripts. LFM Open License v1.0: free for research and for commercial use below the revenue threshold the licence states.
LFM2.5-8B-A1B needs about 6 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.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 3.5 GB | ~272 tok/s | Fits | |
| Q3_K_M 3.41 bpw | 4.3 GB | ~259 tok/s | Fits | |
| Q4_K_M 4.83 bpw | 5.8 GB | ~237 tok/s | Fits | |
| Q5_K_M 5.67 bpw | 6.7 GB | ~226 tok/s | Fits | |
| Q6_K 6.56 bpw | 7.6 GB | ~215 tok/s | Fits | |
| Q8_0 8.50 bpw | 9.6 GB | ~195 tok/s | Fits | |
| F16 16.00 bpw | 17.4 GB | ~143 tok/s | Fits |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. LFM2.5-8B-A1B VRAM calculator →
Get LFM2.5-8B-A1B running
The cheapest catalogued GPU that runs LFM2.5-8B-A1B is the Intel Arc B570 (10 GB).
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.
How to run LFM2.5-8B-A1B
Install Ollama, then run:
ollama run lfm2.5:8bSpecifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 8.3 Billion (1.5B active)
- Context window
- 131,072
- Architecture
- Hybrid LIV convolution + GQA MoE
- Provider
- Liquid AI
- Licence
- LFM Open License v1.0
- Practical deployment
- Laptop
- Specified at
- Q4_K_M
- System RAM
- 16 GB
- Record updated
- 2026-09-11
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Commercial use is licensed only below $10M annual revenue; above that, a separate agreement.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run LFM2.5-8B-A1B on my GPU?
- LFM2.5 on AMD Radeon RX 9060 XT 8GB
- LFM2.5 on Intel Arc B570
- LFM2.5 on Intel Arc B580
- LFM2.5 on NVIDIA GeForce RTX 3060 (12GB)
- LFM2.5 on NVIDIA GeForce RTX 3070
- LFM2.5 on NVIDIA GeForce RTX 3070 Ti
- LFM2.5 on NVIDIA GeForce RTX 3080 (10GB)
- LFM2.5 on NVIDIA GeForce RTX 4060
- LFM2.5 on NVIDIA GeForce RTX 4070
- LFM2.5 on NVIDIA GeForce RTX 4070 Super
- LFM2.5 on NVIDIA GeForce RTX 4070 Ti
- LFM2.5 on NVIDIA GeForce RTX 5060
- LFM2.5 on NVIDIA GeForce RTX 5060 Ti 8GB
- LFM2.5 on NVIDIA GeForce RTX 5070
Other LFM2.5 sizes
LFM2.5-8B-A1B — frequently asked questions
How much VRAM does LFM2.5-8B-A1B need?
About 6 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does LFM2.5-8B-A1B run on an RTX 4090 (24 GB)?
Yes. LFM2.5-8B-A1B needs about 6 GB at Q4_K_M, inside a 24 GB card, at an estimated 237 tokens/sec.
How do I run LFM2.5-8B-A1B locally?
Install Ollama and run `ollama run lfm2.5:8b`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.
What other sizes does LFM2.5 come in?
LFM2.5 2.6B (2 GB), LFM2.5-8B-A1B (6 GB), LFM2.5-24B-A2B (15 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.