The Next Era for Generative Speech

MARS8: The Multilingual Text-to-Speech Model Built for Production

MARS8 is a family of production-grade text-to-speech models built so every use case, language, and voice profile gets the same rock-solid reliability when millions are listening.

Launching natively on all top compute platforms

The World's First Family of TTS Models

The MARS8 family - specialized models for each use-case

01 / 04

MARS8-Flash

Low-latency multilingual TTS for Conversational AI Agents.

Parameters: 600M

USE CASES

  • Real-time voice agents
  • Contact centers
  • Live conversational AI
Try it on our API

02 / 04

MARS8.1-Pro

Highest quality target. Improved pronunciation, expressiveness with high-pitch references, prosody, and accent control / coverage.

Parameters: 600M

USE CASES

  • Expressive dubbing
  • Audiobooks
  • Digital media
Try it on our API

03 / 04

MARS8-Instruct

Fine-grained control over emotion, timing, and style; independent of speaker identity.

Parameters: 1.2B

USE CASES

  • Film & TV dubbing
  • Precise prosody control
  • Creative editing workflows
Try it on our API

04 / 04

MARS8-Nano

When memory and compute are constrained but production quality still matters.

Parameters: 50M

USE CASES

  • Automotive systems
  • Embedded devices
  • Edge deployments
MARS8 FAMILY
The MARS8 familyCAMB.AI feeds MARS8, a family of four specialised models stacked as layers: MARS-Flash, MARS-Pro, MARS-Instruct and MARS-Nano. The active model lifts and lights up while the others separate from it.MARS8MARS-NANOMARS-INSTRUCTMARS-PROMARS-FLASH

NEWMARS8.1 is now in beta. Faster and more expressive, same API.

Read our full research blog post →

Benchmark Results

Redefining the new baseline in TTS

Run the benchmarks yourself →
METRIC

MARS8-Pro

CAMB.AI

MARS8-Flash

CAMB.AI

Sonic-3

Cartesia

Speech-2.6-hd

Minimax

Multilingual_v2

ElevenLabs

Multilingual_v3

ElevenLabs

PQ ↑

Predicted mean opinion score (1–10) from Meta's Audiobox-Aesthetics model; higher indicates better production quality.

7.4498
7.4523
6.9471
6.9468
7.4516
7.1934

WavLM SV cosine similarity ↑

Mean cosine similarity between generated and reference audio using the wavlm-base-sv embedding model.

0.8676
0.8666
0.8420
0.8666
0.8109
0.8253

CAM++ cosine similarity ↑

Mean cosine similarity between generated and reference audio using the CAM++ embedding model.

0.7097
0.7066
0.5134
0.5878
0.3912
0.336

CE ↑

Predicted mean opinion score (1–10) from Meta's Audiobox-Aesthetics model; higher reflects greater content enjoyment.

5.4308
5.4299
5.0445
4.9877
5.4146
5.1816

CER ↓

Percentage of characters incorrect in the generated output, measured by Whisper ASR. Lower is better.

5.77%
5.67%
8.54%
11.30%
4.39%
14.62%

Predicted mean opinion score (1–10) from Meta's Audiobox-Aesthetics model; higher indicates better production quality.

MARS8-Pro

CAMB.AI

7.4498

MARS8-Flash

CAMB.AI

7.4523

Sonic-3

Cartesia

6.9471

Speech-2.6-hd

Minimax

6.9468

Multilingual_v2

ElevenLabs

7.4516

Multilingual_v3

ElevenLabs

7.1934

Languages support for 99% of the world

Global language coverage

MARS8 is the multilingual backbone that lets you cover 99% of the world while staying native to how your audiences speak and listen.

English
Hindi (India)
French (France)
Spanish (Spain)
German
Japanese
Modern Standard Arabic
Korean
Chinese (Simplified)
Italian
Spanish (Mexico)
Portuguese (Portugal)
Portuguese (Brazil)
Indonesian
Dutch

Want the full technical breakdown?

For a detailed look at MARS8's architecture, deployment patterns, and performance characteristics, read the full technical article on our blog.

MARS8 architecture

Build, scale, or partner on your terms

MARS8 is designed to scale across startups, enterprises, and infrastructure providers. Whether you're building a product or enabling others to build, there's a direct path to get started.