Skip to content

LFM2.5 24B A2B on RTX 3090: VRAM, speed and fit

LFM2.5 24B A2B at 6-bit runs on the RTX 3090 with 128K of context at about 204 tokens a second.

Text generation · Chat · Ollama / LM Studio

Runs comfortably · on RTX 3090 · 6-bit (Q6_K) · Ollama / LM Studio · Chat

GPU memory
21.6 GB
of 24.0 GB
Context
128K
auto
Speed
204 tok/s
111 when full
First reply
0.4 s
2K prompt

Where the memory goes

Model
18.3 GB
Context
2.5 GB
Overhead
775 MB
OS
0.0 GB
Free
2.4 GB
Memory
21.6 GB of 24.0 GB
Context
128K tokens · auto
Speed
204 tok/s · 111 when full
First reply
0.4 s for a 2K prompt
Running cost
$6.35/mo at home · $33/mo on a cloud RTX A5000

Validation: Validated

What LFM2.5 24B A2B on RTX 3090 was checked against
CheckTargetResultStatus
Fits in GPU memory24.0 GB usable21.6 GB neededPass
Headroom5% free2.4 GB free (10%)Pass
Context for chat16K or more128KPass
Writing speed15 tok/s or more204 tok/sPass
First replyunder 3 s, 2K prompt0.4 sPass

Targets: a reply you can read along with, a first reply the use case tolerates, and 5% of memory left free. Calculated from published specs, not benchmarked.

The setup

Model
LFM2.5 24B A2B
Parameters
24B, 2B active
Precision
6-bit (Q6_K)
Context
128K, auto
Chats at once
1
Use
Chat
Runs in
Ollama / LM Studio
Hardware
RTX 3090
GPU can use
24.0 GB of 24.0 GB

Running cost

At home
$6.35/mo
290 W at the wall
Renting instead
$33/mo
RTX A5000 at $0.27/h

4 h a day, power at $0.18 per kWh, cloud at RunPod on-demand prices as of Nov 22, 2026.

How to run it

Ollama
ollama run lfm2.5:24b
llama.cpp
llama-server -hf unsloth/LFM2.5-24B-A2B-GGUF
vLLM
vllm serve LiquidAI/LFM2.5-24B-A2B
MLX (Mac)
mlx_lm.chat --model mlx-community/LFM2.5-24B-A2B-OptiQ-4bit

LFM2.5 24B A2B on similar machines

All machines: LFM2.5 24B A2B VRAM requirements. Everything that runs here: what can the RTX 3090 run?

Other models that run on RTX 3090

Questions

Can LFM2.5 24B A2B run on RTX 3090?

LFM2.5 24B A2B at 6-bit runs on the RTX 3090 with 128K of context at about 204 tokens a second.

Report VR-1RIQ0Z8, calculated Nov 22, 2026 from published specs, not benchmarked.