Hardware / Android /Android phone · 6 GB

Android · Phone

Android phone · 6 GB

for example Pixel 7a, Galaxy A54, Redmi Note 13

Runs 49 models fully in memory, 0 more with CPU offload. 5 of those sit close enough to 3 GB that our size estimate is what decides them.

Memory one app may use

3 GB of 6 GB

the most one app may use, not the memory in the phone — estimated

Bandwidth

28 GB/s

Runs

49

01

Specifications

Memory one app may use

3 GB of 6 GB

Bandwidth

28 GB/s

FP16 compute

—

TDP

—

MSRP

—

Street price

—

Released

—

Architecture

—

02

What it runs

Best quant per model — sorted by fit score.

ModelParamsBest quantContextSpeedVerdict

Niagara 19m Batch.en20MQ4_K_M~262K1039 tok/sestFits in memory Niagara 38m Batch.en38MQ4_K_M~262K547 tok/sestFits in memory Parakeet TDT CTC 110m0.1BQ4_K_M~262K189 tok/sestFits in memory Moonshine Streaming Small0.1BQ4_K_M~66K208 tok/sestFits in memory Moonshine Streaming Tiny30MQ5_K_M~262K591 tok/sestFits in memory Moonshine Tiny30MQ5_K_M~262K591 tok/sestFits in memory Canary 180m Flash0.2BQ4_K_M~131K116 tok/sestFits in memory Moonshine Streaming Medium0.2BQ4_K_M~131K87 tok/sestFits in memory Zipformer cr CTC Transducer XL 290M0.3BQ4_K_M~131K72 tok/sestFits in memory Zipformer Transducer XL 290M0.3BQ4_K_M~131K72 tok/sestFits in memory Audio8 ASR 0.1B0.3BQ4_K_M~66K69 tok/sestFits in memory Hubert Large Ls960 ft0.3BQ4_K_M~131K65 tok/sestFits in memory Wav2vec2 Large 960h Lv60 Self0.3BQ4_K_M~131K65 tok/sestFits in memory ASR Conformer Largescaleasr0.5BQ4_K_M~66K43 tok/sestFits in memory Nemotron Speech Streaming en 0.6b0.6BQ4_K_M~66K35 tok/sestFits in memory Parakeet CTC 0.6b0.6BQ4_K_M~66K35 tok/sestFits in memory Parakeet RNNT 0.6b0.6BQ4_K_M~66K35 tok/sestFits in memory Parakeet TDT 0.6b v20.6BQ4_K_M~66K35 tok/sestFits in memory Parakeet TDT 0.6b v30.6BQ4_K_M~66K35 tok/sestFits in memory Nemotron 3.5 ASR Streaming 0.6b0.6BQ4_K_M~66K33 tok/sestFits in memory Distil Large v3.50.8BQ4_K_M~66K26 tok/sestFits in memory Whisper Large v3 Turbo0.8BQ4_K_M~66K26 tok/sestFits in memory Qwen3 ASR 0.6B0.8BQ4_K_M~66K27 tok/sestFits in memory ARK ASR 0.6B1.3BQ4_K_M~66K16 tok/sestFits in memory Owsm CTC v3.1 1B1.1BQ4_K_M~33K19 tok/sestFits in memory Owsm CTC v3.2 ft 1B1BQ4_K_M~33K21 tok/sestFits in memory Owsm CTC v4 1B1BQ4_K_M~33K21 tok/sestFits in memory MMS 1b All1BQ4_K_M~33K21 tok/sestFits in memory Canary 1b1BQ4_K_M~33K21 tok/sestFits in memory Canary 1b Flash1BQ4_K_M~33K21 tok/sestFits in memory Canary 1b v21BQ4_K_M~33K21 tok/sestFits in memory Parakeet CTC 1.1b1.1BQ4_K_M~33K19 tok/sestFits in memory Parakeet RNNT 1.1b1.1BQ4_K_M~33K19 tok/sestFits in memory Parakeet TDT 1.1b1.1BQ4_K_M~33K19 tok/sestFits in memory Llama 3.2 1B Instruct1.2BQ4_K_M33K17 tok/sestFits in memory Lite Whisper Large v3 Acc1.4BQ4_K_M~33K15 tok/sestFits in memory Nyra Health CrisperWhisper1.6BQ4_K_M~33K13 tok/sestFits in memory Whisper Large v31.5BQ4_K_M~33K14 tok/sestFits in memory Qwen3 ASR 0.6B HF0.8BQ4_K_M~8K26 tok/sestFits in memory MOSS Transcribe Diarize0.9BQ4_K_M~8K23 tok/sestFits in memory Qwen3 ASR 1.7B2BQ4_K_M~16K10 tok/sestFits in memory Qwen3 ASR 1.7B HF2BQ4_K_M~16K10 tok/sestFits in memory Cohere Transcribe 03 20262.1BQ4_K_M~16K10 tok/sestFits in memory GLM ASR Nano 25122.3BQ4_K_M~4K9 tok/sestFits in memory Granite 4.0 1b Speech2.3BQ4_K_M~2K9 tok/sestFits in memoryest Granite Speech 4.1 2b2.3BQ4_K_M~2K9 tok/sestFits in memoryest Granite Speech 4.1 2b NAR2.3BQ4_K_M~2K9 tok/sestFits in memoryest STT 2.6b en2.6BQ4_K_M~2K8 tok/sestFits in memoryest Canary Qwen 2.5b2.6BQ4_K_M~2K8 tok/sestFits in memoryest Laguna S 2.1118BQ4_K_M——Too large LongCat 2.01.8TQ8_0——Too large Inkling952BQ5_K_M——Too large Kimi K32.8TQ4_K_M——Too large Hy3299BQ8_0——Too large Nex-N2-Mini35.1BQ4_K_M——Too large GLM 5.2753BQ4_K_M——Too large Kimi K2.7 Code1.1TQ8_0——Too large Nemotron 3 Ultra561BQ5_K_M——Too large MiniMax M3427BQ8_0——Too large Step 3.7 Flash201BQ4_K_M——Too large

A verdict marked est is a close call: the weight size behind it is calculated from the parameter count, and a 10% error would change it.

A speed marked est is calculated from this card's memory bandwidth and the size of the weights it has to read per word. It is not a measurement.

Something wrong on this page? Tell us