Zyphra Technologies Inc.
Palo Alto AI company building compact language, vision-language, and text-to-speech models including the open-source Zamba and Zonos model families.
Zyphra Technologies Inc. is an artificial intelligence research and product company headquartered in Palo Alto, California. The company was founded in 2021 by Krithik Puthalath and Danny Martinelli with a focus on building efficient, small-footprint AI models suitable for on-device and edge deployment.
Zyphra's primary model family is Zamba, a series of hybrid language models combining Mamba state space model (SSM) blocks with global shared attention layers. The original Zamba-7B, released in April 2024, was benchmarked against LLaMA 1, LLaMA 2 7B, and OLMo-7B, claiming superior performance on standard benchmarks while requiring less than half the training data of comparable models. In July 2024, Zyphra launched Zamba2-7B, a follow-on model reporting 25% faster time-to-first-token, 20% improved token generation speed, and lower memory usage compared to competing 7B-class models such as Mistral-7B, Llama3-8B, and Gemma-7B.
In early 2025, Zyphra released Zonos, a pair of open-weight text-to-speech (TTS) models trained on over 200,000 hours of multilingual speech data. The Zonos models are capable of voice cloning from as little as five seconds of audio sample. One Zonos variant uses a pure transformer architecture; the second is a hybrid combining transformer and Mamba SSM architectures. Both models were released under the permissive Apache 2.0 license on Hugging Face.
In 2026, Zyphra released the Zamba2-VL series of vision-language models in 1.2B, 2.7B, and 7B parameter sizes, extending the Zamba2 architecture to multimodal inputs. The company has also been developing MaiaOS, a multimodal agent system intended for autonomous AI applications.
Zyphra has completed a Series A funding round and has reported approximately $8.8 million in annual revenue with a team of around 44 employees as of 2024. The company operates primarily in the AI model research and deployment space, with model weights published on Hugging Face and code on GitHub. While Zyphra's core focus is on language and speech models, the Mamba architecture has been noted by the company as well-suited for real-time robotics inference workloads requiring local on-device execution.
Primary type & automation activities this supplier delivers:
Zamba-7B
- Model: Zamba-7B-v1
- Launch Year: 2024
Open-source 7B hybrid SSM/attention language model outperforming LLaMA 2 7B with less than half the training data.
Zamba2-7B
- Model: Zamba2-7B
- Launch Year: 2024
Efficient 7B small language model with 25% faster TTFT and 20% improved token generation vs Mistral-7B and Llama3-8B.
Zonos
- Model: Zonos-v0.1
- Launch Year: 2025
Open-weight TTS model trained on 200k+ hours of speech, capable of voice cloning from 5 seconds of audio.
More related news
- 1
- 2
- 3
- 4
- 5
-
6
CIIF 2026 | China International Industry Fair | Oct 12–16, Shanghai | CIROS Robot Show RobotToday · May 31, 2026
Contact Zyphra Technologies Inc.
WEBSITE
https://www.zyphra.comHEADQUARTERS
United States
Company Facts
Founded
2021
Primary Role
Software/Algorithm
Company Size
-
Primary Region
North America
Annual Sales
-
Funding Stage
-
Funding Total
-
