QuantDiff

Best LFM2.5 24B A2B quant for your GPU

6 quantized files of LiquidAI/LFM2.5-24B-A2B (23.97B): accuracy, tool calls and speed (no KLD: reference too large).

LiquidAI/LFM2.5-24B-A2B (LFM Open License v1.0). Uploaded to Hugging Face Oct 12, 2026. Mixture of experts: 24.0B total, 2.45B active parameters per token; 40 layers (30 short-convolution blocks + 10 attention), 64 experts with top-4 routing, so the KV cache is small. Always thinks before answering (its chat template opens a think block even when thinking is turned off), so our fixed 1,024-token answer budget can cut off some answers. GGUF quants from unsloth (UD dynamic) and LiquidAI's own.

Accuracy and speed only, no KLD: reference too large. The BF16 reference (about 48 GB) does not fit in our 32 GB test Mac (minipro), so there is no KLD for this model; we measured accuracy, tool calls and speed only. We never estimate KLD; picks for this model are ranked by measured accuracy (GSM8K + MMLU + ARC-C, 500 items, 95% CI) with tool calls as tie-break.

Charts

Every dot and bar is a measured file (GGUF). Hover or tap one for its numbers; each chart has a table under "Show the numbers". Download a PNG to share.

LFM2.5 24B A2B: accuracy vs file size

If a dot's bar reaches the shaded band, that file scores the same as the best one within the margin of error on our accuracy questions (GSM8K + MMLU + ARC-C); the smallest such file here is UD-Q2_K_XL at 7.49 GB.

LFM2.5 24B A2B: accuracy vs file sizeScatter of 6 GGUF files of LFM2.5 24B A2B: file size in GB across, accuracy in percent up, with 95% interval bars. Shaded band: the best file's 95% interval (68.7 to 76.5%). Data table follows.60657075805.07.510.012.515.017.5File size (GB)Accuracy (% of 500 questions)UD-Q3_K_XL by unsloth · 10.20 GB · accuracy 70.0% (65.8–73.9) · tied with the bestQ4_K_M by LiquidAI · 13.53 GB · accuracy 71.4% (67.3–75.2) · tied with the bestUD-Q4_K_XL by unsloth · 13.86 GB · accuracy 70.8% (66.7–74.6) · tied with the bestUD-Q2_K_XL by unsloth · 7.49 GB · accuracy 70.2% (66.0–74.0) · tied with the best · our 12 GB pickQAD-Q4_0 by LiquidAI · 12.65 GB · accuracy 72.2% (68.1–75.9) · tied with the best · our 16 GB pickUD-Q5_K_XL by unsloth · 16.69 GB · accuracy 72.8% (68.7–76.5) · tied with the best · our 24 GB pickUD-Q2_K_XLQAD-Q4_0UD-Q5_K_XLUD-Q3_K_XLUD-Q4_K_XL
  • LiquidAI
  • unsloth
  • best file's 95% range
  • clearly below the best
  • our pick for 8 / 12 / 16 GB GPUs

Up: in %, bars = · Across: file size in GB (GiB) · = overlaps the band

Show the numbers (table)
QuantGBAccuracy %95% intervalCorrectTied with bestPick
UD-Q2_K_XL7.4970.266.0–74.0351/500yes12 GB pick
UD-Q3_K_XL10.2070.065.8–73.9350/500yes
QAD-Q4_012.6572.268.1–75.9361/500yes16 GB pick
Q4_K_M13.5371.467.3–75.2357/500yes
UD-Q4_K_XL13.8670.866.7–74.6354/500yes
UD-Q5_K_XL16.6972.868.7–76.5364/500yes24 GB pick

LFM2.5 24B A2B: tool calls vs file size

If a dot's bar reaches the shaded band, that file scores the same as the best one within the margin of error on our tool-call tests; the smallest such file here is UD-Q2_K_XL at 7.49 GB.

LFM2.5 24B A2B: tool calls vs file sizeScatter of 6 GGUF files of LFM2.5 24B A2B: file size in GB across, tool calls in percent up, with 95% interval bars. Shaded band: the best file's 95% interval (79.4 to 89.3%). Data table follows.7075808590955.07.510.012.515.017.5File size (GB)Tool calls (% of 200 cases)UD-Q3_K_XL by unsloth · 10.20 GB · tool calls 81.0% (75.0–85.8) · tied with the bestQ4_K_M by LiquidAI · 13.53 GB · tool calls 84.5% (78.8–88.9) · tied with the bestUD-Q4_K_XL by unsloth · 13.86 GB · tool calls 83.5% (77.7–88.0) · tied with the bestUD-Q2_K_XL by unsloth · 7.49 GB · tool calls 82.0% (76.1–86.7) · tied with the best · our 12 GB pickQAD-Q4_0 by LiquidAI · 12.65 GB · tool calls 84.0% (78.3–88.4) · tied with the best · our 16 GB pickUD-Q5_K_XL by unsloth · 16.69 GB · tool calls 85.0% (79.4–89.3) · tied with the best · our 24 GB pickUD-Q2_K_XLQAD-Q4_0UD-Q5_K_XLUD-Q3_K_XLQ4_K_MUD-Q4_K_XL
  • LiquidAI
  • unsloth
  • best file's 95% range
  • clearly below the best
  • our pick for 8 / 12 / 16 GB GPUs

Up: in %, bars = · Across: file size in GB (GiB) · = overlaps the band

Show the numbers (table)
QuantGBTool calls %95% intervalCorrectTied with bestPick
UD-Q2_K_XL7.4982.076.1–86.7164/200yes12 GB pick
UD-Q3_K_XL10.2081.075.0–85.8162/200yes
QAD-Q4_012.6584.078.3–88.4168/200yes16 GB pick
Q4_K_M13.5384.578.8–88.9169/200yes
UD-Q4_K_XL13.8683.577.7–88.0167/200yes
UD-Q5_K_XL16.6985.079.4–89.3170/200yes24 GB pick

LFM2.5 24B A2B: diff from the original (KLD)

No KLD for this model: The BF16 reference (about 48 GB) does not fit in our 32 GB test Mac (minipro), so there is no KLD for this model; we measured accuracy, tool calls and speed only.

LFM2.5 24B A2B: generation speed by quant and machine

Longer bar = faster replies. Smaller files are usually faster; the fastest measured here is QAD-Q4_0 (LiquidAI) at 89 tok/s on Apple M2 Pro 32GB (llama.cpp). Compare bars of the same colour only: each colour is a different machine.

LFM2.5 24B A2B: generation speed by quant and machineHorizontal bars of measured generation speed in tokens per second for 6 files on 1 machine setup(s): Apple M2 Pro 32GB (llama.cpp). Longer = faster. Data table follows.0255075100125Generation speed (tokens per second), higher = fasterUD-Q5_K_XLunsloth · 16.7 GBUD-Q5_K_XL by unsloth · 16.69 GB · 62.8 tok/s on Apple M2 Pro 32GB (llama.cpp)63QAD-Q4_0LiquidAI · 12.6 GBQAD-Q4_0 by LiquidAI · 12.65 GB · 88.9 tok/s on Apple M2 Pro 32GB (llama.cpp)89Q4_K_MLiquidAI · 13.5 GBQ4_K_M by LiquidAI · 13.53 GB · 78.5 tok/s on Apple M2 Pro 32GB (llama.cpp)78UD-Q4_K_XLunsloth · 13.9 GBUD-Q4_K_XL by unsloth · 13.86 GB · 69.1 tok/s on Apple M2 Pro 32GB (llama.cpp)69UD-Q3_K_XLunsloth · 10.2 GBUD-Q3_K_XL by unsloth · 10.20 GB · 68.8 tok/s on Apple M2 Pro 32GB (llama.cpp)69UD-Q2_K_XLunsloth · 7.5 GBUD-Q2_K_XL by unsloth · 7.49 GB · 73.8 tok/s on Apple M2 Pro 32GB (llama.cpp)74
  • Apple M2 Pro 32GB (llama.cpp)

Across: (tokens per second)

Show the numbers (table)
FilePublisherGBMachinetok/sPlacement
UD-Q5_K_XLunsloth16.69Apple M2 Pro 32GB (llama.cpp)62.8full GPU
QAD-Q4_0LiquidAI12.65Apple M2 Pro 32GB (llama.cpp)88.9full GPU
Q4_K_MLiquidAI13.53Apple M2 Pro 32GB (llama.cpp)78.5full GPU
UD-Q4_K_XLunsloth13.86Apple M2 Pro 32GB (llama.cpp)69.1full GPU
UD-Q3_K_XLunsloth10.20Apple M2 Pro 32GB (llama.cpp)68.8full GPU
UD-Q2_K_XLunsloth7.49Apple M2 Pro 32GB (llama.cpp)73.8full GPU

Not sure which fits? Use the quant picker →

Recommendation by VRAM

"Best" = highest measured accuracy (500 items pooled) among files that fit fully on the GPU (file + 8192-token KV cache), tool-call score as tie-breaker. No KLD for this model (reference too large).

8GB GPU

Best accuracy that fits

Nothing we've tested fits fully in 8GB at 8192 context. Try a smaller model, run it GPU+RAM split, or rent a bigger GPU.

No GPU big enough? Rent one on RunPod or Vast.ai.

Referral links: QuantDiff may earn credit if you sign up. Referral partners have no say in scores or rankings.

12GB GPU

Best accuracy that fits
UD-Q2_K_XL by unsloth
File size
7.49 GB
no KLD: reference too large
Accuracy (500 items)
70.2% (66.0–74.0)
82.0%
Accuracy
Generation ()
74 t/s on Apple M2 Pro 32GB full GPU
LFM2.5-24B-A2B-UD-Q2_K_XL.gguf
How to run this file

From unsloth/LFM2.5-24B-A2B-GGUF, file LFM2.5-24B-A2B-UD-Q2_K_XL.gguf.

(downloads it, then starts a chat; type /bye to quit):

ollama run hf.co/unsloth/LFM2.5-24B-A2B-GGUF:LFM2.5-24B-A2B-UD-Q2_K_XL.gguf

: open the Discover tab, search unsloth/LFM2.5-24B-A2B-GGUF, pick the UD-Q2_K_XL download, then load it from the Chat tab.

Step-by-step guide

16GB GPU

Best accuracy that fits
QAD-Q4_0 by LiquidAI
File size
12.65 GB
no KLD: reference too large
Accuracy (500 items)
72.2% (68.1–75.9)
84.0%
Accuracy
Generation ()
89 t/s on Apple M2 Pro 32GB full GPU
LFM2.5-24B-A2B-QAD-Q4_0.gguf
How to run this file

From LiquidAI/LFM2.5-24B-A2B-GGUF, file LFM2.5-24B-A2B-QAD-Q4_0.gguf.

(downloads it, then starts a chat; type /bye to quit):

ollama run hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:LFM2.5-24B-A2B-QAD-Q4_0.gguf

: open the Discover tab, search LiquidAI/LFM2.5-24B-A2B-GGUF, pick the QAD-Q4_0 download, then load it from the Chat tab.

Step-by-step guide

All LFM2.5 24B A2B quants

QuantPublisherSizeFits meanKLD p99PP Gen Run
UD-Q2_K_XL unsloth imatrix 7.49 GB 12GB+ n/a n/a — 82.0% 84.7
78–90		56.5 ≈
50–63		74.0 ≈
66–80		 796 74 Apple M2 Pro 32GB settings & logs
UD-Q3_K_XL unsloth imatrix 10.20 GB 12GB+ n/a n/a — 81.0% 76.7 ≈
69–83		59.5 ≈
53–66		77.3 ≈
70–83		 798 69 Apple M2 Pro 32GB settings & logs
QAD-Q4_0 LiquidAI 12.65 GB 16GB+ n/a n/a — 84.0% 76.0 ≈
69–82		61.5 ≈
55–68		82.7 ≈
76–88		 920 89 Apple M2 Pro 32GB settings & logs
Q4_K_M LiquidAI 13.53 GB 16GB+ n/a n/a — 84.5% 76.7 ≈
69–83		59.5 ≈
53–66		82.0 ≈
75–87		 811 78 Apple M2 Pro 32GB settings & logs
UD-Q4_K_XL unsloth imatrix 13.86 GB 16GB+ n/a n/a — 83.5% 76.7 ≈
69–83		61.0 ≈
54–67		78.0 ≈
71–84		 799 69 Apple M2 Pro 32GB settings & logs
UD-Q5_K_XL unsloth imatrix 16.69 GB 24GB+ n/a n/a — 85.0% 76.0 ≈
69–82		62.5
56–69		83.3
77–88		 754 63 Apple M2 Pro 32GB settings & logs

Accuracy cells: % correct, with the 95% confidence interval (Wilson) underneath. ≈ = statistically tied with the best quant of the same model (intervals overlap). ▼ below = clearly worse than that model's best quant (intervals do not overlap). Ties are not ranked: sorting by an accuracy column puts tied quants together and orders them by KLD. How accuracy is measured.

GB = GiB (1024³ bytes). Measured on: tool calls: minipro (Apple M2 Pro 32GB); speed: minipro (Apple M2 Pro 32GB); accuracy: minipro (Metal, Apple M2 Pro (Mac mini), 32 GB unified memory). On a different GPU, expect generation speed to scale roughly with memory bandwidth. Task accuracy (GSM8K / MMLU / ARC-C, % correct on fixed subsets of 150 / 200 / 150 items, frozen suite quantdiff-lmeval-v1) was measured with the same GGUF files and llama.cpp build on Ghost (CUDA, NVIDIA GeForce RTX 5070 12 GB (64 GB RAM)), minipro (Metal, Apple M2 Pro (Mac mini), 12-core CPU, 19-core GPU, 32 GB unified memory), minipro (Metal, Apple M2 Pro (Mac mini), 32 GB unified memory). Accuracy depends on the weights, not on the GPU, but small numeric differences between backends can flip individual answers. With these sample sizes, differences under about 5 points are within noise (± shown on run pages). Methodology.

GPU too small? Rent one.

Run a 32B model at Q6/Q8, or a 70B, on a cloud GPU for an hour or two instead of buying hardware.

RunPod

Cloud GPU pods with 24-80GB cards (RTX 4090 / 5090 / A100 / H100) for big quants or BF16 reference runs.

Rent on RunPod
Vast.ai

GPU marketplace for renting a 24-48GB card for an evening of testing.

Rent on Vast.ai
Your GPU cloud here

Sponsor slot. Sponsors never influence scores or rankings.

Sponsor QuantDiff

Referral links: QuantDiff may earn credit if you sign up. Referral partners have no say in scores or rankings.