whatmodel.sh · find the best speech-to-text models you can run

Best multilingual speech-to-text models to run locally on a 16GB graphics card

A leaderboard of the best multilingual speech-to-text models you can run locally on a 16GB graphics card (with 32GB of system RAM). SeamlessM4T v2 Large is #1 right now, followed by Cohere Transcribe (03-2026) and Whisper large-v3. Models are ranked on transcription tests in many languages. Each model shows the download that fits and the command to run it.

28 models · scores combine 7 benchmarks · updated 2026-10-11 02:20 UTChow it works · find a leaderboard · benchmark your machine

task[all tasks][long-form transcription][multilingual transcription]

Scores use reviewed multilingual transcription evaluations only. Compare scores within this task. Models without task measurements are excluded.

Models for 32GB RAM · 16GB graphics card → ~16GB fast · local only · reset

+ more

card[RTX 4060 Ti 16GB][RTX 4070 Ti Super][RTX 4080][RTX 4080 Super][RTX 5060 Ti 16GB][RTX 5070 Ti][RTX 5080][RTX 3080 Laptop][RTX 3080 Ti Laptop][RTX 4090 Laptop][RTX 5080 Laptop][RTX A4000][RX 7600 XT][RX 7800 XT][RX 7900 GRE][RX 9060 XT][RX 9070][RX 9070 XT]

Detect your exact chip or graphics card:

curl -G whatmodel.sh/stt -d max=0 --data-urlencode "mac=$(sysctl -n hw.memsize machdep.cpu.brand_string)"Mac

curl -G whatmodel.sh/stt -d max=0 --data-urlencode "ram=$(awk '/MemT/{print $2}' /proc/meminfo)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null)"Linux

curl.exe -G whatmodel.sh/stt -d max=0 --data-urlencode "ram=$((gcim Win32_ComputerSystem).TotalPhysicalMemory)" --data-urlencode "nv=$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>$null)"Windows

$ terminal

ram[8][16][24][32][48][64][96][128][192][256]GB

gpu[apple silicon][graphics card][none]

vram[1][2][4][8][12][16][24][32][48][64][80][96][128][160]GB

price[local][$0.001][$0.002][$0.005][any]max per audio min

detecting your OS…

Score vs memory to run

     │                              Cohere Transcribe (03-202…                   ● SeamlessM4T v2 Large
     │   ·                      ·     Whisper large-v3 ●──────●─────∘────────────┘   ·     ∘              ∘ ·
  50 ┤ · · · · · · · · · · · · ·●────────────●─────────∘ · · ·∘· · ·∘· · · · ·∘· · · · ·∘· · · ·∘· ·∘· · · · ·
     │   ·                      │            ∘         ∘      ·                      ·  ∘                   ·
     │   ●──────────────────────┘
     └───┬──────────────────────┬─────────────────────────────┬──────────────────────┬──────────────────────┬─
        1GB                    2GB                           5GB                   10GB                   20GB

#	model	task score↓	where↓	speed↓	RAM↓	get it	sources

#	model	task score↓	where↓	RAM↓	get it
2	SeamlessM4T v2 Large · fp32 on CPU, fp16 on CUDA (seamless_communication m4t_predict) Meta	79±8*	local	~9.2GB	how to run ›
7	Cohere Transcribe (03-2026) · bf16 Cohere	64±9*	local	~5.5GB	how to run ›
8	Whisper large-v3 · fp16 OpenAI	61±6*	local	~4.2GB	how to run ›
10	Omnilingual ASR LLM 7B v2 · bf16 Meta	59±9*	local (slow)	~19GB	how to run ›
12	Omnilingual ASR LLM 1B v2 · bf16 Meta	59±6*	local	~6GB	how to run ›
13	Voxtral Mini 3B (2507) · bf16 Mistral	57±6*	local	~12GB	how to run ›
14	Omnilingual ASR LLM 7B · bf16 Meta	57±9*	local (slow)	~19GB	how to run ›
15	Omnilingual ASR CTC 1B v2 · bf16 Meta	51±6*	local	~2.9GB	how to run ›
16	Whisper large-v3-turbo · fp16 OpenAI	51±9*	local	~2.4GB	how to run ›
17	Cohere Transcribe Arabic (07-2026) · bf16 (transformers >=5.4.0) Cohere	50±9*	local	~4.1GB	how to run ›
18	Parakeet TDT 0.6B v3 · fp32 NVIDIA	49±6	local	~3.5GB	how to run ›
19	Omnilingual ASR CTC 7B v2 · bf16 Meta	49±9*	local (slow)	~16GB	how to run ›
20	Omnilingual ASR LLM 3B v2 · bf16 Meta	48±9*	local	~11GB	how to run ›
21	Phi-4-multimodal-instruct · bf16 Microsoft	48±6*	local	~14GB	how to run ›
22	Qwen3-ASR-1.7B · bf16 Alibaba (Qwen)	48±5	local	~6.1GB	how to run ›
23	Omnilingual ASR CTC 3B v2 · bf16 Meta	47±9*	local	~7.9GB	how to run ›
24	Omnilingual ASR CTC 7B · bf16 Meta	47±9*	local (slow)	~16GB	how to run ›
25	Omnilingual ASR LLM 3B · bf16 Meta	46±9*	local	~11GB	how to run ›
26	Canary-1B-v2 · onnx-fp32 NVIDIA	46±6*	local	~5.2GB	how to run ›
27	Omnilingual ASR LLM 1B · bf16 Meta	44±9*	local	~6GB	how to run ›
29	Omnilingual ASR CTC 3B · bf16 Meta	41±9*	local	~7.9GB	how to run ›
30	Qwen3-ASR-0.6B · bf16 Alibaba (Qwen)	38±7*	local	~2.8GB	how to run ›
31	Voxtral Mini 4B Realtime (2602) · bf16 Mistral	37±4	local	~11GB	how to run ›
33	Omnilingual ASR LLM 300M v2 · bf16 Meta	35±9*	local	~4.4GB	how to run ›
34	Omnilingual ASR CTC 1B · bf16 Meta	33±9*	local	~2.9GB	how to run ›
35	Omnilingual ASR LLM 300M · bf16 Meta	31±9*	local	~4.4GB	how to run ›
36	Omnilingual ASR CTC 300M v2 · bf16 Meta	14±9*	local	~1.3GB	how to run ›
37	Omnilingual ASR CTC 300M · bf16 Meta	12±9*	local	~1.3GB	how to run ›

+ 9 more don't fit your machine or budget · show them

37 of 37 catalogue models listed here · sources and licences

* new or thinly tested: ranked on too few benchmark results · † only the maker’s own numbers so far · ~RAM is estimated; full context may need more memory