LFM2.5 2.6B · 36 GPUs measured first-party · llama.cpp Q4_K_M · Updated October 2026

What GPU Do You Need for LFM2.5 2.6B?

We ran LFM2.5 2.6B (Q4_K_M, llama.cpp) on 36 GPUs, from budget cards to datacenter parts, and measured generation speed, prompt processing, power draw and peak VRAM on every one. The fastest was NVIDIA GeForce RTX 5090 at 566.6 tok/s.

Benchmarked weights: LiquidAI/LFM2.5-2.6B-GGUF

Fastest we measured

NVIDIA GeForce RTX 5090

566.6 tok/s on LFM2.5 2.6B, the ceiling. Measured on our bench. 32GB of VRAM, $1,999 at launch.

Pros

• 566.6 tok/s on LFM2.5 2.6B
• 32GB, clears the LFM2.5 2.6B floor

Cons

• 575W board rating

Best consumer card

NVIDIA GeForce RTX 4090

385.1 tok/s on LFM2.5 2.6B, fastest card you can buy at retail. Measured on our bench. 24GB of VRAM, $1,599 at launch.

Pros

• 385.1 tok/s on LFM2.5 2.6B
• 24GB, clears the LFM2.5 2.6B floor

Cons

• 450W board rating

Cheapest card that runs it

GeForce RTX 4060

135.2 tok/s on LFM2.5 2.6B, lowest launch price that still fits. Measured on our bench. 8GB of VRAM, $299 at launch.

Pros

• 135.2 tok/s on LFM2.5 2.6B
• 8GB, clears the LFM2.5 2.6B floor

Cons

• 115W board rating

Best value

NVIDIA GeForce RTX 3060 Ti

198.8 tok/s on LFM2.5 2.6B, most speed per dollar. Measured on our bench. 8GB of VRAM, $399 at launch. That is 498.2 tok/s per $1,000 of launch price.

Pros

• 198.8 tok/s on LFM2.5 2.6B
• 8GB, clears the LFM2.5 2.6B floor

Cons

• 200W board rating

What the numbers show. Across 36 GPUs measured on our own bench, RTX 5090 is fastest at 567 tok/s. Per dollar of launch price, RTX 5060 Ti gives the most (572.0 tok/s per $1,000). The fastest card with 16GB or less is RTX 5080 at 377 tok/s. The measured peak was ~2GB, so it runs on 6GB cards and up. That floor is a property of the model, so it applies to every GPU, measured or not.

How it compares. H100 80GB HBM3: LFM2.5 2.6B 482.6 tok/s, gemma-3-1b 504.2 (1B), MiniCPM5 2B 428.7 (3B), Qwen2-1.5B 536.5 (2B), Qwen2.5-1.5B 537.2 (2B). 3 of 4 beat LFM2.5 2.6B here.

Cost on a rented GPU. 1M generated tokens of LFM2.5 2.6B: $0.065 on a RTX 3060 ($0.036/hr, 108 min), $0.19 on a RTX 5090 ($0.39/hr, 29 min, 3.0x the cost).

Speed tiers for LFM2.5 2.6B. 30+ tok/s: 36 (RTX 5090, RTX 4090, RTX 5080). 30 tok/s is roughly where replies outpace reading.

Reading your prompt. Before LFM2.5 2.6B writes anything it reads the input: 33248.6 tok/s on the RTX 5090 (0.1s for a 4,000-token prompt), 23302.8 on the RTX 4090 (0.2s), 2865.3 on the T4 (1.4s). Long documents and big code files feel this number more than the generation speed.

VRAM for LFM2.5 2.6B. Measured peak 1.8GB, so 8GB is the smallest common card size; smallest card it ran on: RTX 3070 Ti (8GB). With long context: Q4_K_M 2GB (tested), Q5_K_M 3GB, Q6_K 3GB, Q8_0 4GB, BF16 13GB.

Power on LFM2.5 2.6B. Most efficient: RTX 5090, 142W, 69.5 Wh per 1M generated tokens. Hungriest: RTX 3090 Ti, 222W, 0.17 kWh. At $0.15/kWh: $0.010 per 1M generated tokens.

Our verdict

Fastest on LFM2.5 2.6B: NVIDIA GeForce RTX 5090, 566.6 tok/s. Cheapest consumer card that ran it: GeForce RTX 4060 ($299, 135.2 tok/s). Cheapest to rent per job: NVIDIA GeForce RTX 3060, $0.065 per 1M generated tokens.

FAQ

What GPU do I need to run LFM2.5 2.6B?

About 2GB. Cheapest consumer card that ran it: GeForce RTX 4060 (8GB, 135.2 tok/s).

How fast is LFM2.5 2.6B on the NVIDIA GeForce RTX 5090?

566.6 tok/s at 142W, faster than every datacenter card we ran it on.

How much does it cost to run LFM2.5 2.6B in the cloud?

$0.065 per 1M generated tokens on a NVIDIA GeForce RTX 3060 at $0.036/hr, cheapest of 31 rentable cards we measured.

Can I run LFM2.5 2.6B on a 12GB, 16GB or 24GB card?

It used 1.8GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.

Is the RTX 4090 or the RTX 3090 faster for LFM2.5 2.6B?

The RTX 4090: 385.1 vs 315.0 tok/s, 22% faster on our bench.

Should I buy or rent a RTX 5090 for LFM2.5 2.6B?

Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 10,481x 1M generated tokens of LFM2.5 2.6B. Buy only if you'll run more than that.

How we test

llama.cpp llama-bench at Q4_K_M, 512-token prompt and 128 generated tokens, three runs after a warmup, full GPU offload, with power and VRAM sampled throughout. Token generation is memory-bandwidth-bound, so the ranking tracks bandwidth closely, which makes it a fair guide to cards we haven't run yet.