LLM Self-Hosting Cost Calculator

Will this model run on my GPU/server? Is it cheaper than an API? — Instantly estimate VRAM fit, token speed, and the self-host-vs-API break-even. No keys, no server, no data leaves the browser.

Model newest first

28B densedensectx 262,144Apache-2.0

Quantization

Hardware

Context length 8,192 tok

Concurrent requests 16

GPU rent override ($/hr, blank = default)

Comparison API preset

Comparison API price ($/1M tok)

Result

✅ Fits on one NVIDIA H100 80GB SXM

VRAM

Weights KV cache Overhead Device VRAM limit

Weights 13.9GB + KV 2.1GB (@8,192 tok) + Overhead 1.9GB = 17.9GB / 80GB

· 멀티모달(image-text)

Inference speed

120.6 tok/s

single stream

Serving throughput

1,350.6 tok/s

serving total (batch 16, VRAM headroom max 29)

Self-host vs API

Estimated self-host cost ($2.5/hr)$0.51 / 1M tok

Selected API price$0.8 / 1M tok

Throughput needed to beat the API868.1 tok/s

At this utilization, self-hosting is cheaper. It only holds if you keep the GPU(s) busy.

All results are memory-bandwidth-based approximations. Actual numbers vary with kernel, framework, paging, and quantization method. A real benchmark is recommended before deciding.

Reference · Latest image-gen & speech (STT/TTS) models and API prices

The items below are NOT token-priced, so they don't compare 1:1 with the token calculator above — image-gen APIs plus speech (STT/TTS) self-host models and API prices, compiled for reference. All are public-info approximations.

Image-gen APIs

gpt-image-2OpenAI토큰 단위 · 텍스트 입력 $5 / 이미지 입력 $8 (표준). 이미지당 단가는 OpenAI 계산기 참고

$30 /1M img-out tok

gpt-image-1-miniOpenAI저가형 · 이미지 입력 $2.50 / 텍스트 입력 $2

$8 /1M img-out tok

grok-imagine-imagexAI1K/2K 출력 · 입력 $0.002/img

$0.02 /이미지

Speech STT/TTS — self-host models

Whisper large-v3STT1.55B · MIT · 인코더-디코더 · 100+ 언어 · 정확도 기준선

~3.1GB

Whisper large-v3-turboSTT809M · MIT · 디코더 프루닝 · 훨씬 빠름

~1.6GB

Parakeet TDT 0.6B v3STT600M · CC-BY-4.0 · FastConformer · 다국어(EU) · 타임스탬프

~1.2GB

Canary-1B v2STT978M · CC-BY-4.0 · 25개 EU 언어 · ASR+번역

~2GB

Qwen3-ASR-1.7BSTT1.7B · Apache-2.0 · 52개 언어 · 스트리밍+오프라인 통합

~3.4GB

Moonshine baseSTT61M · MIT · 엣지용 초경량 · 영어

~0.2GB

VoxCPM2TTS2B · Apache-2.0 · 토크나이저-프리 확산 · 30개 언어 · 48kHz · 클로닝

~4GB

Qwen3-TTSTTS1.7B · Apache-2.0 · 3초 클로닝 · 보이스 디자인 · 10개 언어

~3.4GB

Kokoro-82MTTS82M · Apache-2.0 · 초경량·초고속 · 엣지 인기

~0.2GB

ChatterboxTTS500M · MIT · 23+ 언어 · 제로샷 클로닝 · 감정 조절

~1GB

Orpheus 3BTTS3B · Apache-2.0 · Llama 기반 speech-LLM · ~200ms 스트리밍

~6GB

F5-TTSTTS~335M · CC-BY-NC-4.0 · 플로우매칭 DiT · 제로샷 클로닝 · 비상업 라이선스

~0.7GB

Speech STT/TTS — API prices

gpt-4o-transcribeSTTOpenAI · 입력/출력

$2.50 / $10 /1M 오디오 tok

whisper-1STTOpenAI

$0.006 /분

Deepgram Nova-3STTDeepgram · 스트리밍

$0.0077 /분

AssemblyAI Universal-2STTAssemblyAI

$0.15 /시간

ElevenLabs ScribeSTTElevenLabs

$0.22 /시간

Google STT StandardSTTGoogle

$0.016 /분

gpt-4o-mini-ttsTTSOpenAI · 텍스트 입력 / 오디오 출력

$0.60 / $12 /1M tok

ElevenLabs FlashTTSElevenLabs

$50 /1M자

ElevenLabs Multilingual v3TTSElevenLabs

$100 /1M자

Grok Voice (xAI)TTSxAI · STT $0.10/시간 · S2S $0.05/분

$15 /1M자

Cartesia SonicTTSCartesia · 크레딧 기반 환산 [추정]

~$0.038 /분