Community submitted benchmarks for Local AI.

Every result comes from a signed report produced on community hardware. Browse by model, quantisation and chip.

Install & start · macOS / Linux

curl -LsSf https://computearena.ai/install.sh | sh -s launch

Requires curl, tar and Python 3. Review and confirm the installation; the interactive menu opens in this terminal. No sudo required.

Latest runs

  1. Qwen3.5 35B A3BReported as Qwen3.5-35B-A3B-Q8 · qwen35moebasecomputeAustralia
    111.4
    decode
    2,310
    prefill
  2. Qwen3.5-35B-A3B-Q4Reported as Qwen3.5-35B-A3B-Q4 · qwen35moebasecomputeAustralia
    136.6
    decode
    2,329
    prefill
  3. Qwen3.6-27B-Q4Reported as Qwen3.6-27B-Q4 · qwen35basecomputeAustralia
    38.3
    decode
    410
    prefill
  4. Qwen3.5-2B-Base-Q8Reported as Qwen3.5-2B-Base-Q8 · qwen35basecomputeAustralia
    228.6
    decode
    5,433
    prefill
  5. Qwen3.5-2B-Q8Reported as Qwen3.5-2B-Q8 · qwen35basecomputeAustralia
    227.4
    decode
    5,418
    prefill
  6. Qwen3.5-2B-Q4Reported as Qwen3.5-2B-Q4 · qwen35basecomputeAustralia
    313.4
    decode
    5,478
    prefill
How it works
01

Install ComputeArena

Get the ComputeArena CLI for macOS or Linux, then choose BaseRT or llama.cpp as your runtime.

02

Run offline

No account is required. Signed reports remain on your machine until you submit them.

03

Submit when ready

Sign in, review the data in your saved reports, and submit the benchmarks you want to share publicly.

Leaderboard

Model × quantisation × chip

Headline rankings compare compatible PP512 prefill and TG128 decode workloads. The measured workload is shown beside every result.

Model
Runtime
Quantisation
Chip
Prefill size
Rank by
433 matching configurations
#Model / quantRuntimeChip / backendUserUpdatedShare
1Tinystories Lay8 HS512 HD8 33MReported as ivnle/tinystories-lay8-hs512-hd8-33M · llama · default-q4
Q4
BaseRTApple M5 Pro
Metal
2,068.0
TG128
190,107
PP512
arki05Germany
2Tinystories Lay8 HS512 HD8 33MReported as RichardErkhov/ivnle_-_tinystories-lay8-hs512-hd8-33M-gguf · llama
Q4_0
llama.cppApple M5 Pro
BLAS + Metal
1,443.3
TG128
124,181
PP512
arki05Germany
3Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen · default-q4
Q4
BaseRTApple M5 Max
Metal
708.3
TG128
34,136
PP512
basecomputeAustralia
4Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q4
BaseRTApple M4 Max
Metal
583.3
TG128 @ 1 ctx
9,998
PP512
basecomputeAustralia
5Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen · default-q8
Q8
BaseRTApple M5 Max
Metal
549.5
TG128
33,088
PP512
basecomputeAustralia
6Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen · default-q4
Q4
BaseRTApple M5 Pro
Metal
516.8
TG128
20,778
PP512
arki05Germany
7Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q4_K_M
llama.cppAMD Radeon RX 7900 XT (RADV NAVI31)
Vulkan
508.8
TG128
26,137
PP512
arki05Germany
8Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q4
BaseRTApple M3 Ultra
Metal
501.7
TG128 @ 1 ctx
12,830
PP512
basecomputeAustralia
9Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q8_0
llama.cppAMD Radeon RX 7900 XT (RADV NAVI31)
Vulkan
479.4
TG128
26,346
PP512
arki05Germany
10Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q8
BaseRTApple M4 Max
Metal
466.7
TG128 @ 1 ctx
10,006
PP512
basecomputeAustralia
11Qwen3 0.6BReported as Qwen3-0.6B-cuda-q4, Qwen3-0.6B-cuda-q4mix, Qwen3-0.6B-cuda-q8, Qwen/Qwen3-0.6B · qwen
Q4
BaseRTNVIDIA GB10
CUDA
458.8
TG128 @ 1 ctx
52,750
PP512
basecomputeAustralia
12Qwen3 0.6BReported as Qwen3-0.6B-cuda-q4, Qwen3-0.6B-cuda-q4mix · qwen
Q4
BaseRTunknown
CUDA
450.2
TG128 @ 1 ctx
51,746
PP512
basecomputeAustralia
13Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q8
BaseRTApple M3 Ultra
Metal
440.2
TG128 @ 1 ctx
12,494
PP512
basecomputeAustralia
14Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q4_K_M
llama.cppApple M5 Max
BLAS + Metal
439.1
TG128
24,365
PP512
basecomputeAustralia
15Gemma 3 1B ITReported as google/gemma-3-1b-it · gemma3 · default-q4
Q4
BaseRTApple M5 Max
Metal
415.5
TG128
23,019
PP512
lukasAustralia
16Llama 3.2 1B InstructReported as meta-llama/Llama-3.2-1B-Instruct · llama
Q4
BaseRTApple M3 Ultra
Metal
415.1
TG128 @ 1 ctx
9,066
PP512
basecomputeAustralia
17Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q4
BaseRTApple M4 Pro
Metal
401.9
TG128 @ 1 ctx
5,025
PP512
basecomputeAustralia
18Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q8_0
llama.cppApple M5 Max
BLAS + Metal
379.0
TG128
24,983
PP512
basecomputeAustralia
19Llama 3.2 1B InstructReported as meta-llama/Llama-3.2-1B-Instruct · llama
Q4
BaseRTApple M4 Max
Metal
371.7
TG128 @ 1 ctx
6,067
PP512
basecomputeAustralia
20Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q4_K_M
llama.cppAMD Radeon RX 7900 XT
ROCm
368.6
TG128
23,972
PP512
arki05Germany
21Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q4
BaseRTApple M1 Max
Metal
366.4
TG128 @ 1 ctx
5,861
PP512
basecomputeAustralia
22Gemma 3 1B ITReported as google/gemma-3-1b-it · gemma3
Q4
BaseRTApple M4 Max
Metal
361.2
TG128 @ 1 ctx
7,705
PP512
basecomputeAustralia
23Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q4_K_M
llama.cppApple M5 Pro
BLAS + Metal
358.3
TG128
14,552
PP512
arki05Germany
24Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen · default-q8
Q8
BaseRTApple M5 Pro
Metal
350.2
TG128
20,537
PP512
arki05Germany
25Gemma 3 1B ITReported as google/gemma-3-1b-it · gemma3
Q4
BaseRTApple M3 Ultra
Metal
331.4
TG128 @ 1 ctx
10,125
PP512
basecomputeAustralia
26Qwen3 1.7BReported as Qwen/Qwen3-1.7B · qwen
Q4
BaseRTApple M4 Max
Metal
320.1
TG128 @ 1 ctx
4,146
PP512
basecomputeAustralia
27Qwen3 1.7BReported as Qwen/Qwen3-1.7B · qwen
Q4
BaseRTApple M3 Ultra
Metal
314.9
TG128 @ 1 ctx
5,751
PP512
basecomputeAustralia
28Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen3
Q8_0
llama.cppAMD Radeon RX 7900 XT
ROCm
313.7
TG128
24,691
PP512
arki05Germany
29Qwen3.5-2B-Q4Reported as Qwen3.5-2B-Q4 · qwen35
Q4
BaseRTApple M3 Ultra
Metal
313.4
TG128 @ 1 ctx
5,478
PP512
basecomputeAustralia
30Qwen3.5-2B-Base-Q4Reported as Qwen3.5-2B-Base-Q4 · qwen35
Q4
BaseRTApple M3 Ultra
Metal
312.3
TG128 @ 1 ctx
5,495
PP512
basecomputeAustralia
31Qwen3.5 2BReported as Qwen/Qwen3.5-2B · qwen35
Q4
BaseRTApple M3 Ultra
Metal
311.3
TG128 @ 1 ctx
5,484
PP512
basecomputeAustralia
32Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q8
BaseRTApple M1 Max
Metal
307.9
TG128 @ 1 ctx
5,884
PP512
basecomputeAustralia
33Qwen3.5 2B BaseReported as Qwen/Qwen3.5-2B-Base · qwen35
Q4
BaseRTApple M3 Ultra
Metal
305.2
TG128 @ 1 ctx
5,442
PP512
basecomputeAustralia
34Qwen3.5-2B-Q4Reported as Qwen3.5-2B-Q4 · qwen35
Q4
BaseRTApple M4 Max
Metal
303.2
TG128 @ 1 ctx
3,949
PP512
basecomputeAustralia
35Qwen3.5 2BReported as Qwen/Qwen3.5-2B · qwen35
Q4
BaseRTApple M4 Max
Metal
301.7
TG128 @ 1 ctx
1,718
PP512
basecomputeAustralia
36Qwen3.5-2B-Base-Q4Reported as Qwen3.5-2B-Base-Q4 · qwen35
Q4
BaseRTApple M4 Max
Metal
298.8
TG128 @ 1 ctx
3,966
PP512
basecomputeAustralia
37Qwen3 0.6BReported as Qwen/Qwen3-0.6B · qwen
Q8
BaseRTApple M4 Pro
Metal
298.7
TG128 @ 1 ctx
4,845
PP512
basecomputeAustralia
38Gemma 3 1B ITReported as gemma-3-1b-it-cuda-q4mix, gemma-3-1b-it-cuda-q8, google/gemma-3-1b-it · gemma3
Q4
BaseRTNVIDIA GB10
CUDA
295.9
TG128 @ 1 ctx
33,841
PP512
basecomputeAustralia
39Qwen3.5 2B BaseReported as Qwen/Qwen3.5-2B-Base · qwen35
Q4
BaseRTApple M4 Max
Metal
295.1
TG128 @ 1 ctx
1,713
PP512
basecomputeAustralia
40Gemma 3 1B ITReported as google/gemma-3-1b-it · gemma3 · default-q4
Q4
BaseRTApple M5 Pro
Metal
293.1
TG128
15,066
PP512
arki05Germany