Featherless launches fixed-fee GLM 5.2 private cloud

Featherless has launched a private cloud version of the GLM 5.2 AI model for a flat monthly fee of USD $7,500, saying it cuts inference costs by 94% compared with closed-source rivals.

The service is built around Featherless's optimisation of Z.ai's open-source model to run natively on AMD hardware in private cloud infrastructure. That lets customers avoid token-based billing and instead pay a fixed annual cost of USD $90,000 for a fully utilised development team.

According to Featherless, a team using about 100 billion tokens a month would pay USD $1,557,600 a year with GPT-5.5 and USD $1,506,000 with Claude Opus 4.8. Under Featherless's pricing model, the same workload would cost USD $90,000 a year, implying monthly savings of about USD $118,000 to USD $122,000 and annual savings of more than USD $1.46 million.

Featherless is targeting engineering departments that want predictable spending for coding workloads. It argues that variable token charges at large scale have become a constraint for software teams using frontier AI models for sustained development work.

At the centre of the launch is GLM 5.2, a recently released open-source model from Z.ai that Featherless is offering through an OpenAI-compatible API. As a Day Zero launch partner, Featherless hosts the model directly, removing the need for customers to provision their own graphics processing units.

The private cloud version supports context windows of up to 1 million tokens, while the public cloud version supports up to 256,000 tokens. Featherless also offers hosting in Europe and the United States and says it does not keep logs.

AMD focus

A key part of the announcement is the use of AMD rather than Nvidia hardware for native execution of GLM 5.2. Featherless says it is the only platform to have achieved this optimisation for the model, which it argues helps customers avoid supply shortages and high procurement costs linked to Nvidia chips.

The claim reflects a broader shift in the AI infrastructure market, where demand for Nvidia processors has pushed up costs for companies seeking to deploy large models at scale. By building around AMD, Featherless is presenting a lower-cost route for enterprises that want private deployments without relying on usage-based pricing from major model providers.

Eugene Cheah, Chief Executive Officer and Co-Founder of Featherless, outlined the company's view of the economics of AI software development.

"The financial reality of closed-source AI models has become the leading bottleneck in enterprise software scalability. Spending over a million dollars every year on tokens is inherently constraining engineering speed and pushing firms for their efficiency. The open source framework is the way forward in software development since it breaks vendor lock-in and provides unparalleled economics. With GLM 5.2 and AMD optimisation, we are delivering to enterprises complete technological freedom and very clear budgets," said Eugene Cheah, Chief Executive Officer and Co-Founder of Featherless.

Model performance

Featherless says GLM 5.2 is designed for software engineering tasks and long coding sessions. The model uses a Mixture-of-Experts design with 744 billion parameters, of which 39 billion are active per token, and includes changes intended to improve coding and reasoning performance over the previous generation.

Those changes include a 1 million token context window, an IndexShare mechanism designed for long coding sessions, and a revised multi-token prediction layer that Featherless says improves speculative decoding acceptance by about 20%. It also includes built-in controls that let users choose between High and Max settings for reasoning and speed trade-offs.

The model is released under the MIT licence, making it one of the more permissively licensed large models available to businesses seeking open-source alternatives. That may matter for companies looking to reduce dependence on closed systems while retaining flexibility in deployment and integration.

To support its case for GLM 5.2 in software development, Featherless cited a series of benchmark results. It said Terminal-Bench 2.1 scores rose from 63.5 to 81.0 and SWE-bench Pro scores increased from 58.4 to 62.1.

In longer and more complex coding tests, the reported gains were larger. FrontierSWE rose from 30.5 to 74.4, while SWE-Marathon increased from 1.0 to 13.0.

Featherless also pointed to reasoning benchmarks, saying AIME 2026 scores improved from 95.3 to 99.2 and GPQA-Diamond scores rose from 86.2 to 91.2. It says these results place GLM 5.2 among the leading open-source models for long-term programming tasks and closer to the performance of premium closed-source systems.

The launch follows Featherless's partnership with Z.ai to distribute GLM 5.2 globally. For customers, that means access to the model through Featherless's managed infrastructure rather than having to run the architecture themselves, a distinction that may appeal to engineering teams seeking private deployments without building the underlying hosting stack.

Featherless says the model is intended as a drop-in alternative for enterprise software development teams that currently rely on products such as Claude Opus 4.8 and GPT-5.5. Its challenge to those providers rests less on raw benchmark scores than on whether a fixed-fee structure can persuade companies to trade familiar closed systems for an open-source model running on AMD infrastructure.