Smallest.ai launches Lightning V3, claims to beat ElevenLabs on conversational TTS
Smallest.ai released Lightning V3 on March 27, 2026, a text-to-speech model targeting real-time voice agents. The San Francisco-based startup claims the model beats OpenAI, Cartesia, and ElevenLabs on conversational quality benchmarks. The claims are self-reported, and the model weights are not publicly available.
What Lightning V3 is
Lightning V3 is a non-autoregressive TTS model that generates audio at 44,100 Hz. Smallest.ai says it requires under 1 GB of VRAM and achieves a real-time factor of 0.01, meaning it generates 10 seconds of speech in 100 milliseconds. Time-to-first-audio is under 100 milliseconds according to the company (Smallest.ai blog, Mar 27, 2026).
The release includes two sub-versions. V3.1 is the default model optimized for natural conversation, with a reported WVMOS score of 5.06. V3.2 adds instruction-following: it accepts directives for emotion, pitch, volume, and whispering.
Both versions support 15 languages as of March 27, 2026: English, Spanish, French, Italian, Dutch, Swedish, Portuguese, German, Hindi, Tamil, Kannada, Telugu, Malayalam, Marathi, and Gujarati (Smallest.ai blog, Mar 27, 2026). The model handles automatic language detection and mid-sentence code-switching.
Seven of those 15 languages are Indic. That is not a coincidence.
Listen: Lightning V3 vs ElevenLabs
Both samples generated by VoiceRadar on March 30, 2026 via official APIs. Same text for both, only the last sentence differs (provider name). Smallest.ai: Lightning V3.1, voice Magnus, 44,100 Hz, mp3. ElevenLabs: Flash v2.5, voice Roger, 44,100 Hz, mp3. No post-processing applied. Note: Lightning V3.1 mispronounces the ".io" TLD in "voiceradar.io." ElevenLabs Flash v2.5 handles it correctly. The samples are published unmodified.
The company behind it
Smallest.ai was founded in 2023 by Sudarshan Kamath and Akshat Mandloi, both former Bosch engineers and IIT Guwahati graduates. Kamath previously led AI perception for self-driving vehicles at Bosch and contributed to Toppr’s $150 million acquisition (Inc42, Aug 2025).
The company has raised $8.26 million across four rounds, with an $8 million seed led by Sierra Ventures in August 2025. Other investors include 3one4 Capital, Better Capital, AWS Startups, and individual angel Shyamal Hitesh Anadkat, who works at OpenAI (Crunchbase).
Beyond TTS, Smallest.ai operates two products: Waves (TTS, voice cloning, voice conversion) and Atoms (a real-time voice agent platform). The company reported handling over 1 million calls per month for enterprise customers and serving 5,000+ SMBs across India and the United States as of its August 2025 funding announcement (Inc42, Aug 2025).
The benchmark claims
Smallest.ai reports the following scores for Lightning V3 as of March 27, 2026:
- MOS (conversational): 3.89
- WVMOS (V3.1): 5.06
- Intonation: 3.33 (highest among competitors tested)
- Prosody: 3.07 (highest among competitors tested)
- WER: 5.38%
- Preference vs OpenAI gpt-4o-mini-tts: 76.2% in blind listening tests across 1,088 samples in English, Hindi, Spanish, and Tamil
The preference evaluation used an LLM-as-a-Judge framework on the Seed TTS dataset (Smallest.ai blog, Mar 27, 2026).
VoiceRadar has not independently verified these benchmark claims. All scores cited above are self-reported by Smallest.ai. The preference evaluation uses LLM-as-a-Judge rather than human evaluators. No independent third-party benchmark has been published as of March 29, 2026. Press coverage of the launch consists entirely of ANI/PTI wire syndication, with no coverage from independent tech publications such as TechCrunch, VentureBeat, or The Verge.
The closed-source question
The company name is “Smallest” and the pitch centers on efficiency: under 1 GB VRAM, non-autoregressive architecture, runs on small systems. But the model weights are not publicly available. You cannot download Lightning V3, inspect its architecture, measure its actual VRAM usage, or reproduce the benchmark results. Access is API-only, starting at $9 per month.
This stands in direct contrast to Mistral’s Voxtral TTS, which launched one day earlier on March 26, 2026 with open weights on Hugging Face. Voxtral is a 4B-parameter model that can be downloaded, run locally on a Mac with Apple Silicon via MLX, and independently benchmarked by anyone (Mistral AI, Mar 26, 2026). When Mistral claims performance numbers, you can verify them yourself.
When a company claims “smallest and fastest” but does not release the model, the claims remain marketing until proven otherwise.
Pricing and availability
Lightning V3 is available through Smallest.ai’s API on the Pro plan at $9 per month, with usage at approximately $0.025 per 1,000 characters for V3.1 as of March 2026. Enterprise pricing is custom and includes on-premise deployment, 99.99% uptime SLA, and HIPAA/SOC2/GDPR/ISO 27001 compliance (Smallest.ai).
For comparison as of March 2026:
| Provider | Model | Per 1K chars | Open source |
|---|---|---|---|
| OpenAI | gpt-4o-mini-tts | $0.015 | No |
| Mistral | Voxtral TTS | $0.016 | Yes (open weights) |
| Smallest.ai | Lightning V3.1 | $0.025 | No |
| ElevenLabs | Flash Turbo | $0.06 | No |
| ElevenLabs | Multilingual v3 | $0.12 | No |
OpenAI and Mistral are cheaper per character. OpenAI does not offer voice cloning. Mistral’s model launched days ago and has not been widely tested in production. ElevenLabs remains the most expensive per character but has the largest language support (70+), the broadest product suite, and the strongest enterprise traction as of March 2026.
What this tells us about the TTS market in March 2026
Two TTS models launched within 24 hours of each other in the last week of March 2026. One is open-source with verifiable claims. The other is closed-source with self-reported benchmarks and wire-syndicated press coverage.
Both are priced below ElevenLabs. Both target the voice agent market. Both claim to match or beat ElevenLabs on quality.
The TTS market is fragmenting. Pricing pressure is coming from every direction: open-source models that anyone can run, Indian startups operating with lower cost structures, and big tech integrating TTS into existing platforms. For developers evaluating a TTS provider in 2026, the question is shifting from “which one sounds best” to “which claims can I actually verify.”
VoiceRadar will be running independent benchmarks of open-source TTS models on consumer hardware in the coming weeks. Results will be published here.