QuantDiff

Best Spark-X2.6 4B quant for your GPU

10 quantized files of XHToken/Spark-X2.6-4B (4.11B), each diffed against the BF16 original.

XHToken/Spark-X2.6-4B (Apache 2.0). Released Oct 19, 2026 (Hugging Face upload); same architecture as Spark-X2.5-4B with improved post-training. 4.1B parameters, 36 layers (9 full attention, the rest 512-token sliding window), so the KV cache is small. GGUF quants from bartowski (imatrix), plus the maker's own Q4_K_M and mradermacher's i1-Q4_K_M for comparison. Tool calls and accuracy tested with thinking disabled (the chat template defaults to thinking on).

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.

Spark-X2.6 4B: 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 Q3_K_M at 1.86 GB.

Spark-X2.6 4B: accuracy vs file sizeScatter of 10 GGUF files of Spark-X2.6 4B: file size in GB across, accuracy in percent up, with 95% interval bars. Shaded band: the best file's 95% interval (75.0 to 82.2%). Data table follows.40506070809012345File size (GB)Accuracy (% of 500 questions)IQ2_M by bartowski · 1.46 GB · accuracy 49.0% (44.6–53.4) · clearly below the bestIQ3_XXS by bartowski · 1.68 GB · accuracy 67.2% (63.0–71.2) · clearly below the bestQ3_K_M by bartowski · 1.86 GB · accuracy 72.4% (68.3–76.1) · tied with the bestIQ4_XS by bartowski · 2.18 GB · accuracy 75.8% (71.9–79.3) · tied with the besti1-Q4_K_M by mradermacher · 2.42 GB · accuracy 76.8% (72.9–80.3) · tied with the bestQ4_K_M by bartowski · 2.44 GB · accuracy 78.8% (75.0–82.2) · tied with the bestQ5_K_M by bartowski · 2.95 GB · accuracy 78.2% (74.4–81.6) · tied with the bestQ4_K_M by XHToken · 3.30 GB · accuracy 78.6% (74.8–82.0) · tied with the bestQ6_K by bartowski · 3.36 GB · accuracy 77.8% (74.0–81.2) · tied with the bestQ8_0 by bartowski · 4.07 GB · accuracy 78.8% (75.0–82.2) · tied with the best · our 8/12/16/24 GB pickQ8_0IQ2_MIQ3_XXSQ3_K_MIQ4_XSi1-Q4_K_MQ4_K_MQ5_K_MQ4_K_MQ6_K
  • XHToken
  • bartowski
  • mradermacher
  • 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
IQ2_M1.4649.044.6–53.4245/500no (below)
IQ3_XXS1.6867.263.0–71.2336/500no (below)
Q3_K_M1.8672.468.3–76.1362/500yes
IQ4_XS2.1875.871.9–79.3379/500yes
i1-Q4_K_M2.4276.872.9–80.3384/500yes
Q4_K_M (bartowski)2.4478.875.0–82.2394/500yes
Q5_K_M2.9578.274.4–81.6391/500yes
Q4_K_M (XHToken)3.3078.674.8–82.0393/500yes
Q6_K3.3677.874.0–81.2389/500yes
Q8_04.0778.875.0–82.2394/500yes8/12/16/24 GB pick

Spark-X2.6 4B: 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 IQ4_XS at 2.18 GB.

Spark-X2.6 4B: tool calls vs file sizeScatter of 10 GGUF files of Spark-X2.6 4B: file size in GB across, tool calls in percent up, with 95% interval bars. Shaded band: the best file's 95% interval (75.5 to 86.3%). Data table follows.2040608010012345File size (GB)Tool calls (% of 200 cases)IQ2_M by bartowski · 1.46 GB · tool calls 42.5% (35.9–49.4) · clearly below the bestIQ3_XXS by bartowski · 1.68 GB · tool calls 56.5% (49.6–63.2) · clearly below the bestQ3_K_M by bartowski · 1.86 GB · tool calls 64.0% (57.1–70.3) · clearly below the bestIQ4_XS by bartowski · 2.18 GB · tool calls 76.0% (69.6–81.4) · tied with the besti1-Q4_K_M by mradermacher · 2.42 GB · tool calls 74.0% (67.5–79.6) · tied with the bestQ4_K_M by bartowski · 2.44 GB · tool calls 76.5% (70.2–81.8) · tied with the bestQ5_K_M by bartowski · 2.95 GB · tool calls 78.0% (71.8–83.2) · tied with the bestQ4_K_M by XHToken · 3.30 GB · tool calls 81.5% (75.5–86.3) · tied with the bestQ6_K by bartowski · 3.36 GB · tool calls 79.5% (73.4–84.5) · tied with the bestQ8_0 by bartowski · 4.07 GB · tool calls 79.5% (73.4–84.5) · tied with the best · our 8/12/16/24 GB pickQ8_0IQ2_MIQ3_XXSQ3_K_MIQ4_XSi1-Q4_K_MQ4_K_MQ5_K_MQ4_K_MQ6_K
  • XHToken
  • bartowski
  • mradermacher
  • 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
IQ2_M1.4642.535.9–49.485/200no (below)
IQ3_XXS1.6856.549.6–63.2113/200no (below)
Q3_K_M1.8664.057.1–70.3128/200no (below)
IQ4_XS2.1876.069.6–81.4152/200yes
i1-Q4_K_M2.4274.067.5–79.6148/200yes
Q4_K_M (bartowski)2.4476.570.2–81.8153/200yes
Q5_K_M2.9578.071.8–83.2156/200yes
Q4_K_M (XHToken)3.3081.575.5–86.3163/200yes
Q6_K3.3679.573.4–84.5159/200yes
Q8_04.0779.573.4–84.5159/200yes8/12/16/24 GB pick

Spark-X2.6 4B: diff from the original (mean KLD), by quant

Lower = closer to the original. Here the closest file is Q8_0 (bartowski, 0.0027) and the furthest is IQ2_M (bartowski, 1.0014); the scale is logarithmic, so each grid step is 10× more.

Spark-X2.6 4B: diff from the original (mean KLD), by quantHorizontal bars, one per GGUF file (10), sorted by bit level then KLD; bar length = mean KL divergence vs the BF16 original, on a log scale. Lower is closer to the original. Data table follows.0.0010.010.1110Mean KLD vs BF16 (log scale), lower = closer★ Q8_0bartowski · 4.1 GBQ8_0 by bartowski · 4.07 GB · mean KLD 0.0027 · our 8/12/16/24 GB pick0.0027Q6_Kbartowski · 3.4 GBQ6_K by bartowski · 3.36 GB · mean KLD 0.00990.0099Q5_K_Mbartowski · 2.9 GBQ5_K_M by bartowski · 2.95 GB · mean KLD 0.03060.0306Q4_K_MXHToken · 3.3 GBQ4_K_M by XHToken · 3.30 GB · mean KLD 0.08200.0820Q4_K_Mbartowski · 2.4 GBQ4_K_M by bartowski · 2.44 GB · mean KLD 0.08890.0889i1-Q4_K_Mmradermacher · 2.4 GBi1-Q4_K_M by mradermacher · 2.42 GB · mean KLD 0.11680.1168IQ4_XSbartowski · 2.2 GBIQ4_XS by bartowski · 2.18 GB · mean KLD 0.12880.1288Q3_K_Mbartowski · 1.9 GBQ3_K_M by bartowski · 1.86 GB · mean KLD 0.37510.3751IQ3_XXSbartowski · 1.7 GBIQ3_XXS by bartowski · 1.68 GB · mean KLD 0.50900.5090IQ2_Mbartowski · 1.5 GBIQ2_M by bartowski · 1.46 GB · mean KLD 1.00141.0014
  • mean KLD of one file
  • our pick for 8 / 12 / 16 GB GPUs

Across: vs the original (no unit; 0 = identical)

Show the numbers (table)
QuantPublisherGBMean KLDPick
Q8_0bartowski4.070.00278/12/16/24 GB pick
Q6_Kbartowski3.360.0099
Q5_K_Mbartowski2.950.0306
Q4_K_MXHToken3.300.0820
Q4_K_Mbartowski2.440.0889
i1-Q4_K_Mmradermacher2.420.1168
IQ4_XSbartowski2.180.1288
Q3_K_Mbartowski1.860.3751
IQ3_XXSbartowski1.680.5090
IQ2_Mbartowski1.461.0014

Spark-X2.6 4B: generation speed by quant and machine

Longer bar = faster replies. Smaller files are usually faster; the fastest measured here is IQ4_XS (bartowski) at 62 tok/s on Apple M2 Pro 32GB (llama.cpp). Compare bars of the same colour only: each colour is a different machine.

Spark-X2.6 4B: generation speed by quant and machineHorizontal bars of measured generation speed in tokens per second for 10 files on 1 machine setup(s): Apple M2 Pro 32GB (llama.cpp). Longer = faster. Data table follows.020406080Generation speed (tokens per second), higher = fasterQ8_0bartowski · 4.1 GBQ8_0 by bartowski · 4.07 GB · 38.7 tok/s on Apple M2 Pro 32GB (llama.cpp)39Q6_Kbartowski · 3.4 GBQ6_K by bartowski · 3.36 GB · 44.4 tok/s on Apple M2 Pro 32GB (llama.cpp)44Q5_K_Mbartowski · 2.9 GBQ5_K_M by bartowski · 2.95 GB · 44.6 tok/s on Apple M2 Pro 32GB (llama.cpp)45IQ4_XSbartowski · 2.2 GBIQ4_XS by bartowski · 2.18 GB · 61.5 tok/s on Apple M2 Pro 32GB (llama.cpp)62i1-Q4_K_Mmradermacher · 2.4 GBi1-Q4_K_M by mradermacher · 2.42 GB · 56.1 tok/s on Apple M2 Pro 32GB (llama.cpp)56Q4_K_Mbartowski · 2.4 GBQ4_K_M by bartowski · 2.44 GB · 55.5 tok/s on Apple M2 Pro 32GB (llama.cpp)55Q4_K_MXHToken · 3.3 GBQ4_K_M by XHToken · 3.30 GB · 52.3 tok/s on Apple M2 Pro 32GB (llama.cpp)52IQ3_XXSbartowski · 1.7 GBIQ3_XXS by bartowski · 1.68 GB · 55.2 tok/s on Apple M2 Pro 32GB (llama.cpp)55Q3_K_Mbartowski · 1.9 GBQ3_K_M by bartowski · 1.86 GB · 49.2 tok/s on Apple M2 Pro 32GB (llama.cpp)49IQ2_Mbartowski · 1.5 GBIQ2_M by bartowski · 1.46 GB · 48.9 tok/s on Apple M2 Pro 32GB (llama.cpp)49
  • Apple M2 Pro 32GB (llama.cpp)

Across: (tokens per second)

Show the numbers (table)
FilePublisherGBMachinetok/sPlacement
Q8_0bartowski4.07Apple M2 Pro 32GB (llama.cpp)38.7full GPU
Q6_Kbartowski3.36Apple M2 Pro 32GB (llama.cpp)44.4full GPU
Q5_K_Mbartowski2.95Apple M2 Pro 32GB (llama.cpp)44.6full GPU
IQ4_XSbartowski2.18Apple M2 Pro 32GB (llama.cpp)61.5full GPU
i1-Q4_K_Mmradermacher2.42Apple M2 Pro 32GB (llama.cpp)56.1full GPU
Q4_K_Mbartowski2.44Apple M2 Pro 32GB (llama.cpp)55.5full GPU
Q4_K_MXHToken3.30Apple M2 Pro 32GB (llama.cpp)52.3full GPU
IQ3_XXSbartowski1.68Apple M2 Pro 32GB (llama.cpp)55.2full GPU
Q3_K_Mbartowski1.86Apple M2 Pro 32GB (llama.cpp)49.2full GPU
IQ2_Mbartowski1.46Apple M2 Pro 32GB (llama.cpp)48.9full GPU

Not sure which fits? Use the quant picker →

Recommendation by VRAM

"Best" = smallest diff (lowest mean KLD) among files that fit fully on the GPU (file + 8192-token KV cache), with tool-call score as tie-breaker.

8GB GPU

Smallest diff that fits
Q8_0 by bartowski
File size
4.07 GB
(mean)
0.0027
98.0%
79.5%
Accuracy
Generation ()
39 t/s on Apple M2 Pro 32GB full GPU
Spark-X2.6-4B-Q8_0.gguf
How to run this file

From bartowski/Spark-X2.6-4B-GGUF, file Spark-X2.6-4B-Q8_0.gguf.

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

ollama run hf.co/bartowski/Spark-X2.6-4B-GGUF:Spark-X2.6-4B-Q8_0.gguf

: open the Discover tab, search bartowski/Spark-X2.6-4B-GGUF, pick the Q8_0 download, then load it from the Chat tab.

Step-by-step guide

Fastest that fits: IQ4_XS (bartowski, 62 t/s, KLD 0.1288)

12GB GPU

Smallest diff that fits
Q8_0 by bartowski
File size
4.07 GB
(mean)
0.0027
98.0%
79.5%
Accuracy
Generation ()
39 t/s on Apple M2 Pro 32GB full GPU
Spark-X2.6-4B-Q8_0.gguf
How to run this file

From bartowski/Spark-X2.6-4B-GGUF, file Spark-X2.6-4B-Q8_0.gguf.

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

ollama run hf.co/bartowski/Spark-X2.6-4B-GGUF:Spark-X2.6-4B-Q8_0.gguf

: open the Discover tab, search bartowski/Spark-X2.6-4B-GGUF, pick the Q8_0 download, then load it from the Chat tab.

Step-by-step guide

Fastest that fits: IQ4_XS (bartowski, 62 t/s, KLD 0.1288)

16GB GPU

Smallest diff that fits
Q8_0 by bartowski
File size
4.07 GB
(mean)
0.0027
98.0%
79.5%
Accuracy
Generation ()
39 t/s on Apple M2 Pro 32GB full GPU
Spark-X2.6-4B-Q8_0.gguf
How to run this file

From bartowski/Spark-X2.6-4B-GGUF, file Spark-X2.6-4B-Q8_0.gguf.

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

ollama run hf.co/bartowski/Spark-X2.6-4B-GGUF:Spark-X2.6-4B-Q8_0.gguf

: open the Discover tab, search bartowski/Spark-X2.6-4B-GGUF, pick the Q8_0 download, then load it from the Chat tab.

Step-by-step guide

Fastest that fits: IQ4_XS (bartowski, 62 t/s, KLD 0.1288)

All Spark-X2.6 4B quants

QuantPublisherSizeFits meanKLD p99PP Gen Run
Q8_0 bartowski imatrix 4.07 GB 8GB+ 0.0027 0.032 98.0% 79.5% 92.0
87–95		63.0
56–69		86.7 ≈
80–91		 654 39 Apple M2 Pro 32GB settings & logs
Q6_K bartowski imatrix 3.36 GB 8GB+ 0.0099 0.109 95.7% 79.5% 90.0 ≈
84–94		62.0 ≈
55–68		86.7 ≈
80–91		 589 44 Apple M2 Pro 32GB settings & logs
Q5_K_M bartowski imatrix 2.95 GB 8GB+ 0.0306 0.326 92.6% 78.0% 90.0 ≈
84–94		63.0
56–69		86.7 ≈
80–91		 535 45 Apple M2 Pro 32GB settings & logs
Q4_K_M XHToken 3.30 GB 8GB+ 0.0820 0.856 88.3% 81.5% 90.0 ≈
84–94		62.5 ≈
56–69		88.7 ≈
83–93		 587 52 Apple M2 Pro 32GB settings & logs
Q4_K_M bartowski imatrix 2.44 GB 8GB+ 0.0889 0.847 87.4% 76.5% 90.0 ≈
84–94		62.5 ≈
56–69		89.3
83–93		 579 55 Apple M2 Pro 32GB settings & logs
i1-Q4_K_M mradermacher imatrix 2.42 GB 8GB+ 0.1168 1.157 85.9% 74.0% 90.0 ≈
84–94		59.5 ≈
53–66		86.7 ≈
80–91		 581 56 Apple M2 Pro 32GB settings & logs
IQ4_XS bartowski imatrix 2.18 GB 8GB+ 0.1288 1.172 84.6% 76.0% 89.3 ≈
83–93		59.5 ≈
53–66		84.0 ≈
77–89		 590 62 Apple M2 Pro 32GB settings & logs
Q3_K_M bartowski imatrix 1.86 GB 8GB+ 0.3751 2.920 74.6% 64.0% 84.7 ≈
78–90		55.0 ≈
48–62		83.3 ≈
77–88		 555 49 Apple M2 Pro 32GB settings & logs
IQ3_XXS bartowski imatrix 1.68 GB 8GB+ 0.5090 3.684 70.7% 56.5% 71.3 ▼ below
64–78		56.0 ≈
49–63		78.0 ≈
71–84		 570 55 Apple M2 Pro 32GB settings & logs
IQ2_M bartowski imatrix 1.46 GB 8GB+ 1.0014 5.603 58.9% 42.5% 36.0 ▼ below
29–44		45.0 ▼ below
38–52		67.3 ▼ below
59–74		 574 49 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: KLD: minipro (Apple M2 Pro 32GB); 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.