LFM2 · Liquid AI
Fine-tuning LFM2 24B-A2B
What does it take to fine-tune LFM2 24B-A2B?
LFM2 24B-A2B needs 48 GB for half-precision LoRA training and 28 GB in four-bit, and 32 GB to serve. It supports supervised fine-tuning, under the LFM Open licence. The cheapest qualifying class is A40 at $0.42 per GPU-hour.
The largest LFM2, with 2.3 billion active parameters out of twenty-four billion. Thirty-two gigabytes to serve a 24B model is unusually low, and it is the strongest argument in the catalogue for the mixture-of-experts trade when serving cost dominates.
What to know before choosing it
Thirty-two gigabytes to serve a 24B model is unusually low, and it is the strongest argument in the catalogue for the mixture-of-experts trade when serving cost dominates. Large-tier knowledge, mid-tier serving footprint.
Training is the harder half. Forty-eight gigabytes for half precision, and the usual mixture-of-experts caveat applies: routing can become unbalanced during training in a way that looks like the model simply plateauing.
Specification
| Parameters | 24B2.3B active per token |
|---|---|
| Architecture | Mixture of experts |
| Size tier | Large |
| LoRA training memory | 48 GBHalf precision, frozen base |
| QLoRA training memory | 28 GBFour-bit base, higher-precision adapter |
| Serving memory | 32 GBHalf precision, before attention cache |
| Objectives | SFT |
| Licence | LFM Open |
| Repository | LiquidAI/LFM2-24B-A2B |
What it costs to train
Every class with enough memory for four-bit training, cheapest first. Total cost is the rate multiplied by wall time, so the cheapest rate is not always the cheapest run — a faster class that finishes sooner frequently wins.
| GPU class | VRAM | Training / hr | Serving / hr | Fits |
|---|---|---|---|---|
| A40 | 48 GB | $0.42 | $0.48 | LoRA and QLoRA |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | LoRA and QLoRA |
| A6000 | 48 GB | $0.65 | $0.75 | LoRA and QLoRA |
| L40S | 48 GB | $0.78 | $0.90 | LoRA and QLoRA |
| A100 40 GB | 40 GB | $1.17 | $1.35 | QLoRA only |
| A100 80 GB | 80 GB | $1.68 | $1.94 | LoRA and QLoRA |
| H100 80 GB | 80 GB | $2.59 | $2.99 | LoRA and QLoRA |
| H200 | 141 GB | $4.55 | $5.25 | LoRA and QLoRA |
Serving fits on A40 at $0.48 per hour — before the attention cache, which grows with context length and concurrency.
Good starting point for
- Large-model quality at mid-tier serving cost
- High request volumes against a broad knowledge base
This is a mixture-of-experts model
24B total parameters with 2.3B active per token. Memory scales with the total; inference compute scales with the active count. That makes it attractive when memory is cheaper than compute for your workload, and unattractive when the reverse holds. Quoting only one of the two numbers is how these models get misrepresented in both directions.
Other sizes in this family
| Model | Params | QLoRA | Objectives |
|---|---|---|---|
| LFM2 350M | 350M | 2 GB | SFT |
| LFM2 700M | 700M | 3 GB | SFT |
| LFM2 1.2B | 1.2B | 4 GB | SFT |
| LFM2 2.6B | 2.6B | 6 GB | SFT |
| LFM2 8B-A1B | 8.3B | 12 GB | SFT |
Comparable sizes elsewhere
Frequently asked questions
How much VRAM does it take to fine-tune LFM2 24B-A2B?
48 GB for half-precision LoRA and 28 GB for four-bit QLoRA. Serving needs 32 GB. Preference tuning roughly doubles the training figure, because a frozen reference model is held alongside the one being trained.
What is the cheapest way to fine-tune LFM2 24B-A2B?
Four-bit QLoRA on A40 at $0.42 per GPU-hour is the cheapest class that meets the 28 GB threshold. Note that quantised training is slower per step, so a faster class sometimes costs less over the whole run.
Can I download the weights after fine-tuning LFM2 24B-A2B?
Yes. Every finished run exposes its trained weights for download, and publishing to a model hub is a single call with a generated model card recording the base model and version the adapter applies to.
What licence does LFM2 24B-A2B carry?
LFM Open. The licence follows the fine-tune — a derivative inherits the base model’s terms, and those terms pass to anyone you give the model to.
Last verified 6 August 2026. Memory thresholds are the platform's own admission limits.
Fine-tune LFM2 24B-A2B
Upload a dataset, forecast the run, and see the cost before any compute is leased.