Pipette Leaderboard | On-Device Model Rankings

Leaderboard for
your device

Pareto Frontier

Find models that maximize quality at any given speed or size.

Which Models Are Fast Without Cutting Corners?

The Pareto frontier highlights the strongest quality-speed tradeoffs. Each model filtered to on using at . Dashed lines connect a model’s quant variants.Device:

Responsive & Accurate

Higher

Lower quantization

Same model across quants

Pareto frontier

Per-Model Performance Across the Board

Compare each model across the selected quality, speed, and memory metrics. Bar lengths show relative rank within each metric across all filtered models; the best result fills the lane. Filtered to on using · .Device:

Model config	Selected metrics
q4_k_m LFM2.5-230M	Average Intelligence 18.2 % E2E 346 ms
q4_k_m LFM2.5-350M	Average Intelligence 23.3 % E2E 420 ms
q4_k_m granite-4.0-350m	Average Intelligence 12.7 % E2E 631 ms
q4_k_m LFM2.5-1.2B-Thinking	Average Intelligence 44.2 % E2E 659 ms
q4_k_m LFM2.5-1.2B-Instruct	Average Intelligence 38.2 % E2E 662 ms
q4_k_m Llama-3.2-1B-Instruct	Average Intelligence 18.3 % E2E 705 ms
q4_k_m MiniCPM5-1B	Average Intelligence 37.0 % E2E 718 ms
q4_k_m Qwen3.5-0.8B	Average Intelligence 12.9 % E2E 849 ms
q4_k_m granite-4.0-h-350m	Average Intelligence 14.0 % E2E 880 ms
q4_k_m LFM2.5-8B-A1B	Average Intelligence 56.9 % E2E 1.0k ms
q4_k_m LFM3-1.7B	Average Intelligence 79.5 % E2E 1.1k ms
q4_k_m Qwen3.5-2B	Average Intelligence 36.5 % E2E 1.2k ms
q4_k_m LFM2.5-2.6B	Average Intelligence 70.8 % E2E 1.3k ms
q4_k_m Spark-X2.6-1.7B	Average Intelligence 60.4 % E2E 1.3k ms
q4_k_m MiniCPM5-2B	Average Intelligence 64.7 % E2E 1.3k ms

Best for this metric

How Each Size Tier Stacks Up

Each model’s Average Intelligence Score (%) is plotted against its size tier (X). Dot size grows with the tier. Color = model family, shade = quant precision. Filtered to on using .Device:

Higher

Lower quantization

Same model across quants

Speed & Memory

How fast models run and how much memory they need across context lengths.

Who Comes Out on Top?

Each bar’s height is the model’s Decode Throughput, ranked best first.Models are atonusingatDevice:

Tokens per second (tok/s)

1,0008006004002000

LFM2.5-230M

LFM2.5-350M

granite-4.0-350m

LFM2.5-1.2B-Thinking

LFM2.5-1.2B-Instruct

Llama-3.2-1B-Instruct

MiniCPM5-1B

Qwen3.5-0.8B

granite-4.0-h-350m

LFM3-1.7B

LFM2.5-8B-A1B

Spark-X2.6-1.7B

Qwen3.5-2B

LFM2.5-2.6B

MiniCPM5-2B

Llama-3.2-3B-Instruct

LFM2.5-24B-A2B

granite-4.0-h-1b

Ling-3.0-tiny

Ministral-3-3B-Instruct-2512

ai9stars_G9v3-3B

gemma-4-E2B-it

granite-4.1-3b

granite-4.2-3b

ERNIE-4.5-21B-A3B-Thinking

gpt-oss-20b

North-Mini-Code-1.0

Spark-X2.6-4B

Qwen3.5-4B

granite-4.0-h-micro

Laguna-XS-2.1

MiniCPM-V-4.7-35B-A3B

Nemotron-3-Nano-4B

Qwen3.5-35B-A3B

Qwen3.6-35B-A3B

Nemotron-3-Nano-30B-A3B

Nemotron-3.5-Lightning-30B-A3B

Nemotron-Cascade-2-30B-A3B

gemma-4-26B-A4B-it

Nemotron-3.5-Bolt-E4B

gemma-4-E4B-it

Olmo-3-7B-Think

Nanbeige4.2-3B

GLM-4.7-Flash

Ministral-3-8B-Instruct-2512

granite-4.1-8b

Ornith-1.0-9B

Qwen3.5-9B

Ornith-1.5-9B

Falcon-H1R-7B

Bonsai-27B

Nemotron-Nano-9B-v2

gemma-4-12b-it

Qwen3.5-27B

Qwen3.6-27B

granite-4.2-30b

56 models ranked by Decode Throughput

LFM2.5-230M

q4_k_m

882 tok/s

LFM2.5-350M

q4_k_m

732 tok/s

granite-4.0-350m

q4_k_m

485 tok/s

LFM2.5-1.2B-Thinking

q4_k_m

452 tok/s

LFM2.5-1.2B-Instruct

q4_k_m

449 tok/s

Llama-3.2-1B-Instruct

q4_k_m

421 tok/s

MiniCPM5-1B

q4_k_m

399 tok/s

Qwen3.5-0.8B

q4_k_m

325 tok/s

granite-4.0-h-350m

q4_k_m

313 tok/s

LFM3-1.7B

q4_k_m

304 tok/s

LFM2.5-8B-A1B

q4_k_m

294 tok/s

Spark-X2.6-1.7B

q4_k_m

268 tok/s

Qwen3.5-2B

q4_k_m

245 tok/s

LFM2.5-2.6B

q4_k_m

230 tok/s

MiniCPM5-2B

q4_k_m

218 tok/s

Llama-3.2-3B-Instruct

q4_k_m

196 tok/s

LFM2.5-24B-A2B

q4_k_m

187 tok/s

granite-4.0-h-1b

q4_k_m

184 tok/s

Ling-3.0-tiny

q4_k_m

181 tok/s

Ministral-3-3B-Instruct-2512

q4_k_m

180 tok/s

ai9stars_G9v3-3B

q4_k_m

171 tok/s

gemma-4-E2B-it

q4_k_m

169 tok/s

granite-4.1-3b

q4_k_m

166 tok/s

granite-4.2-3b

q4_k_m

166 tok/s

ERNIE-4.5-21B-A3B-Thinking

q4_k_m

160 tok/s

gpt-oss-20b

q4_k_m

158 tok/s

North-Mini-Code-1.0

q4_k_m

143 tok/s

Spark-X2.6-4B

q4_k_m

137 tok/s

Qwen3.5-4B

q4_k_m

131 tok/s

granite-4.0-h-micro

q4_k_m

131 tok/s

Laguna-XS-2.1

q4_k_m

128 tok/s

MiniCPM-V-4.7-35B-A3B

q4_k_m

127 tok/s

Nemotron-3-Nano-4B

q4_k_m

127 tok/s

Qwen3.5-35B-A3B

q4_k_m

120 tok/s

Qwen3.6-35B-A3B

q4_k_m

120 tok/s

Nemotron-3-Nano-30B-A3B

q4_k_m

117 tok/s

Nemotron-3.5-Lightning-30B-A3B

q4_k_m

116 tok/s

Nemotron-Cascade-2-30B-A3B

q4_k_m

116 tok/s

gemma-4-26B-A4B-it

q4_k_m

107 tok/s

Nemotron-3.5-Bolt-E4B

q4_k_m

107 tok/s

gemma-4-E4B-it

q4_k_m

107 tok/s

Olmo-3-7B-Think

q4_k_m

99 tok/s

Nanbeige4.2-3B

q4_k_m

97 tok/s

GLM-4.7-Flash

q4_k_m

93 tok/s

Ministral-3-8B-Instruct-2512

q4_k_m

90 tok/s

granite-4.1-8b

q4_k_m

88 tok/s

Ornith-1.0-9B

q4_k_m

86 tok/s

Qwen3.5-9B

q4_k_m

86 tok/s

Ornith-1.5-9B

q4_k_m

83 tok/s

Falcon-H1R-7B

q4_k_m

68 tok/s

Bonsai-27B

q1_0

67 tok/s

Nemotron-Nano-9B-v2

q4_k_m

64 tok/s

gemma-4-12b-it

q4_k_m

60 tok/s

Qwen3.5-27B

iq1_m

38 tok/s

Qwen3.6-27B

iq1_m

37 tok/s

granite-4.2-30b

q4_k_m

29 tok/s

Higher

Lower quantization

How Performance Scales With Context Length

Compare throughput, time, and Peak RAM across measured context lengths. Each line is one model + quant combo, filtered to on using .Device:

Prefill Throughput (tok/s)

Decode Throughput (tok/s)

Higher

Lower quantization

The Memory Cost of Throughput

Prefill and Decode Throughput (tok/s) vs. Peak RAM usage (GB, log scale). Showing on using at . Color = model family, shade = quant precision. Decode uses 100 output tokens.Device:

Higher

Lower quantization

PrefillDecodeDot size = model size

Quant Economics

The tradeoffs of reducing model precision: accuracy loss vs. faster inference and lower memory use.

The Accuracy Hit From Quantization

Bar height = Average Intelligence score (%) for each model at each Quantization level on using . Showing matching models by current sort, 2+ quant metrics only.Device:

Color shows model family; shade shows quant precision, with darker shades indicating less compression. The leftmost darkest bar is the full-precision baseline when available. Lines connect each model's quant levels. Larger vertical drops indicate more accuracy lost from compression, while flatter paths indicate stronger quantization resilience. Use the metric selector to change the benchmark and the sort control to choose which quantization leads the ranking. Leaderboard Guide

Average Intelligence (%)

01836547290

Higher

Lower quantization

trend

Pareto Frontier

Find models that maximize quality at any given speed or size.

Leaderboard for
your device

Pareto Frontier

Find models that maximize quality at any given speed or size.

Which Models Are Fast Without Cutting Corners?

The Pareto frontier highlights the strongest quality-speed tradeoffs. Each model filtered to on using at . Dashed lines connect a model’s quant variants.Device:

Responsive & Accurate

Higher

Lower quantization

Same model across quants

Pareto frontier

Per-Model Performance Across the Board

Model config	Selected metrics
q4_k_m LFM2.5-230M	Average Intelligence 18.2 % E2E 346 ms
q4_k_m LFM2.5-350M	Average Intelligence 23.3 % E2E 420 ms
q4_k_m granite-4.0-350m	Average Intelligence 12.7 % E2E 631 ms
q4_k_m LFM2.5-1.2B-Thinking	Average Intelligence 44.2 % E2E 659 ms
q4_k_m LFM2.5-1.2B-Instruct	Average Intelligence 38.2 % E2E 662 ms
q4_k_m Llama-3.2-1B-Instruct	Average Intelligence 18.3 % E2E 705 ms
q4_k_m MiniCPM5-1B	Average Intelligence 37.0 % E2E 718 ms
q4_k_m Qwen3.5-0.8B	Average Intelligence 12.9 % E2E 849 ms
q4_k_m granite-4.0-h-350m	Average Intelligence 14.0 % E2E 880 ms
q4_k_m LFM2.5-8B-A1B	Average Intelligence 56.9 % E2E 1.0k ms
q4_k_m LFM3-1.7B	Average Intelligence 79.5 % E2E 1.1k ms
q4_k_m Qwen3.5-2B	Average Intelligence 36.5 % E2E 1.2k ms
q4_k_m LFM2.5-2.6B	Average Intelligence 70.8 % E2E 1.3k ms
q4_k_m Spark-X2.6-1.7B	Average Intelligence 60.4 % E2E 1.3k ms
q4_k_m MiniCPM5-2B	Average Intelligence 64.7 % E2E 1.3k ms

Best for this metric

How Each Size Tier Stacks Up

Each model’s Average Intelligence Score (%) is plotted against its size tier (X). Dot size grows with the tier. Color = model family, shade = quant precision. Filtered to on using .Device:

Higher

Lower quantization

Same model across quants

Speed & Memory

How fast models run and how much memory they need across context lengths.

Who Comes Out on Top?

Each bar’s height is the model’s Decode Throughput, ranked best first.Models are atonusingatDevice:

Tokens per second (tok/s)

1,0008006004002000

LFM2.5-230M

LFM2.5-350M

granite-4.0-350m

LFM2.5-1.2B-Thinking

LFM2.5-1.2B-Instruct

Llama-3.2-1B-Instruct

MiniCPM5-1B

Qwen3.5-0.8B

granite-4.0-h-350m

LFM3-1.7B

LFM2.5-8B-A1B

Spark-X2.6-1.7B

Qwen3.5-2B

LFM2.5-2.6B

MiniCPM5-2B

Llama-3.2-3B-Instruct

LFM2.5-24B-A2B

granite-4.0-h-1b

Ling-3.0-tiny

Ministral-3-3B-Instruct-2512

ai9stars_G9v3-3B

gemma-4-E2B-it

granite-4.1-3b

granite-4.2-3b

ERNIE-4.5-21B-A3B-Thinking

gpt-oss-20b

North-Mini-Code-1.0

Spark-X2.6-4B

Qwen3.5-4B

granite-4.0-h-micro

Laguna-XS-2.1

MiniCPM-V-4.7-35B-A3B

Nemotron-3-Nano-4B

Qwen3.5-35B-A3B

Qwen3.6-35B-A3B

Nemotron-3-Nano-30B-A3B

Nemotron-3.5-Lightning-30B-A3B

Nemotron-Cascade-2-30B-A3B

gemma-4-26B-A4B-it

Nemotron-3.5-Bolt-E4B

gemma-4-E4B-it

Olmo-3-7B-Think

Nanbeige4.2-3B

GLM-4.7-Flash

Ministral-3-8B-Instruct-2512

granite-4.1-8b

Ornith-1.0-9B

Qwen3.5-9B

Ornith-1.5-9B

Falcon-H1R-7B

Bonsai-27B

Nemotron-Nano-9B-v2

gemma-4-12b-it

Qwen3.5-27B

Qwen3.6-27B

granite-4.2-30b

56 models ranked by Decode Throughput

LFM2.5-230M

q4_k_m

882 tok/s

LFM2.5-350M

q4_k_m

732 tok/s

granite-4.0-350m

q4_k_m

485 tok/s

LFM2.5-1.2B-Thinking

q4_k_m

452 tok/s

LFM2.5-1.2B-Instruct

q4_k_m

449 tok/s

Llama-3.2-1B-Instruct

q4_k_m

421 tok/s

MiniCPM5-1B

q4_k_m

399 tok/s

Qwen3.5-0.8B

q4_k_m

325 tok/s

granite-4.0-h-350m

q4_k_m

313 tok/s

LFM3-1.7B

q4_k_m

304 tok/s

LFM2.5-8B-A1B

q4_k_m

294 tok/s

Spark-X2.6-1.7B

q4_k_m

268 tok/s

Qwen3.5-2B

q4_k_m

245 tok/s

LFM2.5-2.6B

q4_k_m

230 tok/s

MiniCPM5-2B

q4_k_m

218 tok/s

Llama-3.2-3B-Instruct

q4_k_m

196 tok/s

LFM2.5-24B-A2B

q4_k_m

187 tok/s

granite-4.0-h-1b

q4_k_m

184 tok/s

Ling-3.0-tiny

q4_k_m

181 tok/s

Ministral-3-3B-Instruct-2512

q4_k_m

180 tok/s

ai9stars_G9v3-3B

q4_k_m

171 tok/s

gemma-4-E2B-it

q4_k_m

169 tok/s

granite-4.1-3b

q4_k_m

166 tok/s

granite-4.2-3b

q4_k_m

166 tok/s

ERNIE-4.5-21B-A3B-Thinking

q4_k_m

160 tok/s

gpt-oss-20b

q4_k_m

158 tok/s

North-Mini-Code-1.0

q4_k_m

143 tok/s

Spark-X2.6-4B

q4_k_m

137 tok/s

Qwen3.5-4B

q4_k_m

131 tok/s

granite-4.0-h-micro

q4_k_m

131 tok/s

Laguna-XS-2.1

q4_k_m

128 tok/s

MiniCPM-V-4.7-35B-A3B

q4_k_m

127 tok/s

Nemotron-3-Nano-4B

q4_k_m

127 tok/s

Qwen3.5-35B-A3B

q4_k_m

120 tok/s

Qwen3.6-35B-A3B

q4_k_m

120 tok/s

Nemotron-3-Nano-30B-A3B

q4_k_m

117 tok/s

Nemotron-3.5-Lightning-30B-A3B

q4_k_m

116 tok/s

Nemotron-Cascade-2-30B-A3B

q4_k_m

116 tok/s

gemma-4-26B-A4B-it

q4_k_m

107 tok/s

Nemotron-3.5-Bolt-E4B

q4_k_m

107 tok/s

gemma-4-E4B-it

q4_k_m

107 tok/s

Olmo-3-7B-Think

q4_k_m

99 tok/s

Nanbeige4.2-3B

q4_k_m

97 tok/s

GLM-4.7-Flash

q4_k_m

93 tok/s

Ministral-3-8B-Instruct-2512

q4_k_m

90 tok/s

granite-4.1-8b

q4_k_m

88 tok/s

Ornith-1.0-9B

q4_k_m

86 tok/s

Qwen3.5-9B

q4_k_m

86 tok/s

Ornith-1.5-9B

q4_k_m

83 tok/s

Falcon-H1R-7B

q4_k_m

68 tok/s

Bonsai-27B

q1_0

67 tok/s

Nemotron-Nano-9B-v2

q4_k_m

64 tok/s

gemma-4-12b-it

q4_k_m

60 tok/s

Qwen3.5-27B

iq1_m

38 tok/s

Qwen3.6-27B

iq1_m

37 tok/s

granite-4.2-30b

q4_k_m

29 tok/s

Higher

Lower quantization

How Performance Scales With Context Length

Compare throughput, time, and Peak RAM across measured context lengths. Each line is one model + quant combo, filtered to on using .Device:

Prefill Throughput (tok/s)

Decode Throughput (tok/s)

Higher

Lower quantization

The Memory Cost of Throughput

Prefill and Decode Throughput (tok/s) vs. Peak RAM usage (GB, log scale). Showing on using at . Color = model family, shade = quant precision. Decode uses 100 output tokens.Device:

Higher

Lower quantization

PrefillDecodeDot size = model size

Quant Economics

The tradeoffs of reducing model precision: accuracy loss vs. faster inference and lower memory use.

The Accuracy Hit From Quantization

Bar height = Average Intelligence score (%) for each model at each Quantization level on using . Showing matching models by current sort, 2+ quant metrics only.Device:

Average Intelligence (%)

01836547290

Higher

Lower quantization

trend

Pareto Frontier

Find models that maximize quality at any given speed or size.

On-device model benchmark leaderboard

Models benchmarked

Devices benchmarked

Leaderboard foryour device

Pareto Frontier

Which Models Are Fast Without Cutting Corners?

Per-Model Performance Across the Board

How Each Size Tier Stacks Up

Speed & Memory

Who Comes Out on Top?

How Performance Scales With Context Length

Prefill Throughput (tok/s)

Decode Throughput (tok/s)

The Memory Cost of Throughput

Quant Economics

The Accuracy Hit From Quantization

Pareto Frontier

Leaderboard foryour device

Pareto Frontier

Which Models Are Fast Without Cutting Corners?

Per-Model Performance Across the Board

How Each Size Tier Stacks Up

Speed & Memory

Who Comes Out on Top?

How Performance Scales With Context Length

Prefill Throughput (tok/s)

Decode Throughput (tok/s)

The Memory Cost of Throughput

Quant Economics

The Accuracy Hit From Quantization

Pareto Frontier

Leaderboard for
your device

Leaderboard for
your device