Zyphra

Zyphra is an American artificial intelligence company headquartered in San Francisco, California, founded in 2021 by Krithik Puthalath, Beren Millidge, Tomás Figliolia, and Danny Martinelli. The company develops small and mid-sized multimodal models with an emphasis on next-generation neural-network architectures, parameter sharing, and continual learning, and is building Maia, a general-purpose superagent for knowledge-work productivity. Zyphra reached unicorn status in June 2025 with a $100 million Series A at a $1 billion post-money valuation led by Jaan Tallinn, the early-stage AI investor whose prior Series A leads include DeepMind and Anthropic.

At a glance

• Founded: 2021 in San Francisco, California, by Krithik Puthalath, Beren Millidge, Tomás Figliolia, and Danny Martinelli.
• Status: Private. Series A in June 2025 at $1 billion post-money valuation.
• Funding: Approximately $100 million-plus cumulative private capital. Series A of $100 million in June 2025 led by Jaan Tallinn at $1 billion post-money valuation. Earlier seed and bridge financing from undisclosed investors.
• CEO: Krithik Puthalath, Co-Founder and Chief Executive Officer.
• Other notable leadership: Beren Millidge (Co-Founder and Chief Scientist), Tomás Figliolia (Co-Founder and Head of AI Model Architecture), Danny Martinelli (Co-Founder).
• Open weights: Yes. The Zamba and Zamba2 model families have been released open-weights with technical reports.
• Flagship outputs: Zamba2-2.7B and the broader Zamba2 family of small efficient language models; the Maia general-purpose superagent (in development); the AMD-and-IBM-anchored training infrastructure announced in 2025.

Origins

Zyphra was founded in 2021 by Krithik Puthalath, Beren Millidge, Tomás Figliolia, and Danny Martinelli. Millidge had a prior research background in computational neuroscience and predictive-coding architectures; Figliolia brought a model-architecture and systems background; Martinelli and Puthalath rounded out the founding team across architecture and operations. The founding research thesis was oriented around next-generation neural-network architectures, long-term memory, reinforcement learning, and continual learning, with a stated ambition to build general agents that learn across deployment rather than freezing capability at training time.

The 2021 to 2024 period was comparatively low-profile by frontier-AI standards, with the company building research output around the Zamba and Zamba2 small-language-model families. Zamba2-2.7B and adjacent releases were positioned as competitive small language models in the 2 to 3 billion parameter dense class, with technical claims around parameter-sharing efficiency that delivered meaningful inference-throughput improvements relative to peer dense models at similar parameter counts. The release strategy was open-weights with technical reports, putting Zyphra inside the open-weights research community alongside Mistral, Qwen, and DeepSeek peers at smaller parameter scales.

The June 2025 Series A of $100 million at a $1 billion post-money valuation, led by Jaan Tallinn, was the company's most consequential public-facing transition. Tallinn's prior history of leading Series A rounds for DeepMind and Anthropic gave the round an unusual provenance signal; the post-money valuation put Zyphra into the unicorn cohort despite a comparatively quiet public-research surface relative to peer labs at the same valuation.

In October 2025 Zyphra announced a multi-year collaboration with IBM Cloud and AMD to build large-scale AI training infrastructure for the Maia general-purpose superagent, with public framing positioning the partnership as the first full-stack AMD-anchored AI training platform on IBM Cloud. The 2025 to 2026 period continued model-family development and Maia-related infrastructure build.

Mission and strategy

Zyphra's stated mission is to build artificial superintelligence with an emphasis on architectures that improve through continued use and learning across deployment rather than freezing at the end of training. The strategy combines three threads. First, foundation-model research with the Zamba family providing the public-research surface and the parameter-sharing architecture providing the distinctive technical claim. Second, large-scale training infrastructure with IBM Cloud and AMD anchoring the compute layer. Third, the Maia general-purpose superagent product layer targeted at knowledge-work productivity.

The competitive premise reflects Zyphra's positioning as one of the post-frontier insurgent labs alongside Essential AI, Reka AI, Liquid AI, and Sakana AI, with the parameter-sharing architecture and continual-learning research thesis providing distinctive technical positioning relative to peers.

The AMD-and-IBM infrastructure relationship is also strategically distinctive: most frontier and insurgent training has run on Nvidia hardware, and the Zyphra commitment to AMD compute gives the company a differentiated supply-chain posture and arguably a structural cost advantage if AMD's MI300X-and-successor lineup delivers on its claimed price-performance.

Models and products

• Zamba2 family. Open-weights small efficient language models including Zamba2-2.7B and adjacent releases. Technical claims around parameter-sharing efficiency for inference throughput.
• Maia. General-purpose superagent in development; positioned as a knowledge-work-productivity product targeting the enterprise segment.
• AMD-and-IBM training infrastructure. Announced October 2025; multi-year partnership for large-scale model training on AMD GPU infrastructure hosted on IBM Cloud.

Distribution channels are the Hugging Face open-weights distribution surface for Zamba models, IBM Cloud and AMD partner distribution, and direct enterprise relationships through the Maia product surface as the agent product becomes commercially available.

Benchmarks and standing

Zyphra's evaluation framework focuses on parameter-efficiency and inference-throughput metrics for the Zamba family rather than horizontal frontier-model leaderboards. Zamba2-2.7B has been positioned in technical coverage as a competitive entry in the 2 to 3 billion parameter dense class with the parameter-sharing architecture providing the distinctive efficiency claim. The Maia agent has not yet been benchmarked publicly against peer agent products from frontier and insurgent labs.

Industry coverage has characterized Zyphra as a watchable insurgent lab with distinctive technical positioning around parameter-sharing architectures and the Tallinn-led Series A providing institutional-investor visibility. The AMD-and-IBM infrastructure announcement positioned the company in the comparatively narrow set of AI labs committed to non-Nvidia training infrastructure at scale.

Leadership

As of May 2026, Zyphra's senior leadership includes:

• Krithik Puthalath, Co-Founder and Chief Executive Officer.
• Beren Millidge, Co-Founder and Chief Scientist. Computational neuroscience and predictive-coding research background.
• Tomás Figliolia, Co-Founder and Head of AI Model Architecture.
• Danny Martinelli, Co-Founder.
• Senior research and engineering leadership across the model-architecture, training-infrastructure, and Maia-product programs.

Continued senior research recruitment has supported the Zamba model program and the AMD-and-IBM infrastructure partnership.

Funding and backers

• Seed and bridge (2021 to 2024): Earlier capital from undisclosed investors.
• Series A (June 2025): $100 million at $1 billion post-money valuation led by Jaan Tallinn.

Cumulative disclosed private capital approximately $100 million-plus. Tallinn's prior track record as Series A lead for DeepMind and Anthropic provides distinctive provenance: few individual investors carry an equivalent two-out-of-two record on leading frontier-AI Series A rounds at the unicorn-or-better outcome, and the Tallinn lead was widely characterized in industry coverage as a substantive provenance signal for Zyphra. The October 2025 IBM-and-AMD partnership, while not a fundraising round, brought additional strategic-relationship value via committed compute capacity and operating partnership commitments that supplement the disclosed capital base.

Industry position

Zyphra occupies a position as one of the post-frontier insurgent foundation-model labs with distinctive technical positioning around parameter-sharing architectures and continual-learning research, the AMD-and-IBM infrastructure partnership providing a non-Nvidia compute posture, and the Tallinn-led Series A unicorn round providing institutional-investor visibility. Industry coverage has consistently characterized Zyphra as a watchable insurgent lab, though public-research and public-product visibility has been comparatively narrow relative to peer labs at the same valuation tier.

The structural risks are two. First, the small-and-efficient-model segment that the Zamba family targets is crowded: Mistral, Microsoft Phi, Qwen, DeepSeek, and Arcee AFM all compete for the same per-dollar-inference-economics wins, and the differentiation between parameter-sharing architectures and peer dense or sparse-mixture-of-experts approaches at similar parameter counts is narrow on most workloads. Second, the Maia superagent product layer competes against well-resourced incumbent and insurgent agent products from frontier labs and a long tail of agent-native startups; converting research output into a product moat is the principal commercial question.

Competitive landscape

• Mistral AI, Cohere, AI21 Labs. Insurgent foundation-model peers with different model-size and enterprise-positioning.
• Essential AI, Reka AI, Liquid AI, Sakana AI. Post-frontier insurgent peers with similar founder-research positioning.
• Arcee AI, Microsoft Phi. Small-and-efficient-model peers competing on per-dollar inference economics.
• Adept AI, Cognition AI, Imbue, General Reasoning. Agent-product peers; Maia targets adjacent productivity workflows.
• OpenAI, Anthropic, Google DeepMind. Frontier-lab alternatives whose agent products provide direct competitive comparison once Maia ships.

Outlook

• The Maia general-purpose superagent product launch and how the agent surfaces against frontier-lab and insurgent-agent peers.
• Continued Zamba-family releases and the differentiation of the parameter-sharing architecture against peer small-and-efficient-model labs.
• The execution of the AMD-and-IBM training infrastructure partnership and whether the non-Nvidia compute posture delivers measurable cost or scale advantages.
• The Series B or adjacent fundraising timeline given the comparatively narrow Series A relative to peer post-frontier insurgents.
• Continued senior research-talent recruitment against the post-frontier insurgent cohort.

Sources

• Zyphra official site. Company reference.
• IBM and AMD collaborate with Zyphra. October 2025 infrastructure partnership announcement.
• Zyphra blog. Research and product reference.
• Beren Millidge LinkedIn. Co-Founder and Chief Scientist reference.
• Tomás Figliolia LinkedIn. Co-Founder reference.