A French AI startup named ZML has launched new software called ZML/LLMD. This software helps run large language models on many different types of AI chips. The goal is to speed up AI inference regardless of the hardware used.
ZML/LLMD is not open source, but the company is offering it for free at launch. They want to learn how users will use the product before deciding on paid versions later. The software supports popular chips from Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc.
Steeve Morin, ZML’s founder, says the idea is to give people the power to build their own AI systems. He adds that they want to achieve real efficiency gains so AI can spread faster. Morin also mentioned that ZML is now co-designing silicon alongside its software.
In July 2026, ZML raised $20 million from several venture firms. Investors include 20VC, Kima Ventures, LocalGlobe, and Puzzle Ventures. This funding will help ZML improve its AI inference server and expand its reach across hardware types.
Breaking AI Chip Silos
ZML/LLMD aims to break the barriers between AI chips. Many AI tools only work well on one brand of hardware. ZML wants to change that by enabling maximum speed on any chip. This approach could reduce the need for companies to stick to a single chip provider.
This is important because AI models are growing larger and need more computing power. Running these models efficiently on various chips can lower costs and increase accessibility. Morin’s vision is to help AI spread by making it easier for people to use their own systems.
Wider AI Industry Moves
Other AI players have made notable moves recently. For example, GLM-5.2, a massive language model with 744 billion parameters, was released as open-source weights on June 16, 2026. It was trained on 28.5 trillion tokens and runs on Huawei silicon, avoiding American chips entirely.
GLM-5.2 ranks second on Code Arena, trailing only Anthropic’s Claude Fable 5. It offers API pricing that cuts costs by up to 82 percent compared to Anthropic’s Claude Opus 4.8. The model cost about $25 million to train, with 80 percent spent on post-training.
Meanwhile, SambaNova, an AI chip maker, raised $1 billion at an $11 billion valuation in July 2026. CEO Rodrigo Liang highlighted the importance of their deal with JPMorgan Chase, which chose SambaNova for its inference infrastructure. Liang said this sends a message to the banking industry to reduce dependence on cloud services.
SambaNova has a multi-year partnership with Intel and recently unveiled its SN50 chip. This chip will start shipping in the second half of 2026. The company is open to going public and expects more investors to join their Series F round soon.
On June 22, Zhipu’s market capitalization crossed HK$1 trillion, or about US$128 billion. Following the launch of GLM-5.2, JPMorgan raised its revenue forecast for Zhipu by 7 to 16 percent. This shows strong confidence in AI companies pushing model and chip innovation globally.
All these developments show the AI industry is rapidly evolving. Startups like ZML focus on hardware flexibility and efficiency. At the same time, chip makers and model developers race to deliver more powerful and cost-effective AI solutions.
Based on
- Hot French startup ZML releases free product to speed inference across lots of AI chips — techcrunch.com
- A cheap Chinese model is catching up with U.S. AI giants on their home turf – The Japan Times — japantimes.co.jp
- Z.ai launches ZCode to challenge Cursor, Claude Code and GitHub Copilot in AI coding | VentureBeat — venturebeat.com
- OpenAI to launch new model after US freeze — france24.com
- AI chip maker SambaNova raises $1B at $11B valuation, 5 months after last mega round | TechCrunch — techcrunch.com
