Z

Zyphra

Open-source AI company building multimodal agents

Research Engineer - Language Model Pre-Training

Full-Time
No salary listed
Mid
Master's
San Francisco, CA, USA
In Person
AI & Machine LearningAI Research

About the job

Requirements
  • Strong engineering aptitude for rapidly implementing reliable and robust systems.
  • Ability to rapidly learn new fields and implement new ideas.
  • Excellent communication and collaboration skills, with the ability to work effectively on research and engineering implementation at scale.
  • Deep expertise and intuition for solving machine learning problems and training models.
  • Experience training models on large-scale, multi-node graphics processing unit clusters.
  • Deep understanding of model training pipelines, including model parallelism, data parallelism, and distributed optimizers.
  • Strong grasp of proper experimental methodology for running rigorous ablations and other hypothesis testing.
  • Understanding of large-scale, highly parallel data processing pipelines.
  • High proficiency with PyTorch and Python.
  • Ability to dive into large pre-existing codebases and rapidly get up to speed.
  • A postgraduate degree in a scientific subject such as Computer Science, Electrical Engineering, Mathematics, or Physics.
Responsibilities
  • Shape the language model roadmap through end-to-end pretraining development.
  • Work closely with the pretraining team to integrate insights into next-generation models.
  • Conduct large-scale training runs and model parallelization.
  • Optimize the performance of the pretraining stack.
  • Collect, process, and evaluate datasets.
  • Conduct architecture and methodology research, including optimizer ablations.
Desired Qualifications
  • Published machine learning research in well-respected venues.

About the company

Z

Zyphra

View
Website

Zyphra builds open-source, open-science AI focused on multimodal models and efficient systems that run on a wide range of hardware. It develops autonomous agent platforms for enterprises to enable conversational AI, automation, and offline-personalized assistants. Its Maia project blends neural architectures with long-term memory, reinforcement learning, and continual learning for text and audio modalities. Zyphra also releases open-source assets like Zamba, Zamba2-2.7B, and Zyda data, and is backed by investors to support offline, hardware-flexible AI with transparent research resources.

Company Size

51-200

Company Stage

Series B

Total Funding

$621.4M

Headquarters

San Francisco, California

Founded

2019

Get referred to Zyphra

See people who can refer or advise you

Simplify Jobs

Simplify's Take

What believers are saying

  • IBM updated its Zyphra collaboration on August 11, 2026, reinforcing infrastructure access.
  • ZAYA1-8B launched May 6, 2026 on Hugging Face and Zyphra Cloud.
  • ZONOS2 shipped June 12, 2026, expanding Zyphra's open-source audio stack and distribution.

What critics are saying

  • Open-source rivals like DeepSeek, Mistral, and Qwen erase Zyphra's model advantage quickly.
  • AMD and IBM concentration exposes Zyphra to pricing, capacity, and roadmap shocks.
  • If Maia misses enterprise traction in 2026, Zyphra becomes a research lab.

What makes Zyphra unique

  • Zyphra trains open-weight models on AMD-only stacks, reducing NVIDIA dependence and supply constraints.
  • ZAYA1-8B used under 1 billion active parameters to rival much larger reasoning models.
  • Zyphra Cloud bundles inference, post-training, and agent infrastructure for enterprise deployment.

Help us improve and share your feedback! Did you find this helpful?

Benefits

  • Health Insurance
  • Dental Insurance
  • Vision Insurance
  • 401(k) Retirement Plan
  • 401(k) Company Match
  • Relocation Assistance
  • Wellness Program

Growth & Insights and Company News

6 month growth

2%

1 year growth

8%

2 year growth

27%
PR Newswire
May 6th, 2026
Zyphra releases ZAYA1-8B reasoning model with under 1B active parameters, rivalling models many times larger

Zyphra has released ZAYA1-8B, a mixture-of-experts language model using fewer than one billion active parameters that matches or exceeds substantially larger open-weight models on reasoning, mathematics and coding tasks. The model was trained entirely on AMD Instinct MI300X clusters with AMD Pensando Polara networking on IBM Cloud infrastructure. ZAYA1-8B performs competitively with models many times its size across mathematics benchmarks, coding and reasoning tasks, matching models like Nemotron-3-Nano-30B-A3B and Mistral-Small-4-119B whilst remaining competitive with frontier reasoning models including DeepSeek-R1-0528. The company also introduced Markovian RSA, a test-time compute methodology enabling unbounded reasoning whilst keeping memory costs constant. ZAYA1-8B is available as a free serverless endpoint on Zyphra Cloud and on Hugging Face under an Apache 2.0 licence.

PR Newswire
May 4th, 2026
Zyphra launches AI cloud platform on AMD MI355X GPUs with TensorWave

Zyphra has launched Zyphra Cloud, a full-stack AI platform powered by AMD Instinct MI355X GPUs on TensorWave infrastructure. The platform debuts with Zyphra Inference, a serverless inference service for frontier open-weight models including DeepSeek V3.2, Kimi K2.6 and GLM 5.1. The San Francisco-based company combines custom kernels and novel long-context inference algorithms to deliver high-throughput performance for production use cases such as agentic coding and workflow automation. Zyphra Cloud will expand to include distributed post-training services, sandboxed agent environments and dedicated GPU clusters. The platform is available immediately, providing developers and enterprises with unified model serving, agent infrastructure and scalable compute for building advanced AI systems.

PR Newswire
Feb 18th, 2026
Zyphra releases ZUNA, open-source brain-computer interface model for thought-to-text AI

Zyphra has released ZUNA, a foundation model for brain-computer interfaces that processes electroencephalography data and advances towards thought-to-text communication. The 380-million-parameter diffusion autoencoder model reconstructs high-fidelity brain signals from imperfect EEG data, improving diagnostics and research workflows. ZUNA works across various EEG systems, from consumer headsets to 256-electrode research equipment, predicting missing channels from sparse inputs. The model outperforms traditional spherical-spline interpolation methods, particularly with incomplete or noisy data. The San Francisco-based company released ZUNA as open-source software under an Apache 2.0 licence, with model weights available on Hugging Face and code on GitHub. Zyphra is seeking collaborations to improve future versions for specific use cases across medical devices, neuroscience research and consumer neurotechnology sectors.

Zonebourse
Oct 1st, 2025
Zyphra partners with IBM, AMD: $1B AI boost

IBM and AMD have partnered with Zyphra, an AI startup valued at $1 billion after its Series A funding, to provide next-generation AI infrastructure. The multi-year agreement includes deploying a large cluster of AMD Instinct MI300X GPUs and AMD Pensando network accelerators on IBM Cloud. The first phase was delivered in September, with expansion planned for 2026. Zyphra will use this to train its 'Maia' super-agent for enhancing business productivity through language, image, and sound processing.

VentureBeat
Jun 7th, 2024
Zyphra Debuts Zyda, A 1.3T Language Modeling Dataset It Claims Outperforms Pile, C4, Arxiv

VB Transform 2024 returns this July! Over 400 enterprise leaders will gather in San Francisco from July 9-11 to dive into the advancement of GenAI strategies and engaging in thought-provoking discussions within the community. Find out how you can attend here. Zyphra Technologies is announcing the release of Zyda, a massive dataset designed to train language models. It consists of 1.3 trillion tokens and is a filtered and deduplicated mashup of existing premium open datasets, specifically RefinedWeb, Starcoder, C4, Pile, Slimpajama, pe2so, and arxiv. The company claims its ablation studies reveal that Zyda performs better than the datasets it was built on. An early dataset version powers Zyphra’s Zamba model and will eventually be available for download on Hugging Face.Image credit: Zyphra“[We] came up with Zyda when [we] were trying to create a pretraining dataset for [our] Zamba series of models,” Zyphra Chief Executive Krithik Puthalath tells VentureBeat in an email