Featherless has launched a personal cloud model of the GLM 5.2 AI mannequin for a flat month-to-month payment of USD $7,500, saying it cuts inference prices by 94% in contrast with closed-source rivals.
The service is constructed round Featherless’s optimisation of Z.ai’s open-source mannequin to run natively on AMD {hardware} in personal cloud infrastructure. That lets prospects keep away from token-based billing and as an alternative pay a set annual price of USD $90,000 for a totally utilised improvement crew.
In response to Featherless, a crew utilizing about 100 billion tokens a month would pay USD $1,557,600 a 12 months with GPT-5.5 and USD $1,506,000 with Claude Opus 4.8. Below Featherless’s pricing mannequin, the identical workload would price USD $90,000 a 12 months, implying month-to-month financial savings of about USD $118,000 to USD $122,000 and annual financial savings of greater than USD $1.46 million.
Featherless is focusing on engineering departments that need predictable spending for coding workloads. It argues that variable token fees at massive scale have change into a constraint for software program groups utilizing frontier AI fashions for sustained improvement work.
On the centre of the launch is GLM 5.2, a just lately launched open-source mannequin from Z.ai that Featherless is providing by an OpenAI-compatible API. As a Day Zero launch companion, Featherless hosts the mannequin straight, eradicating the necessity for purchasers to provision their very own graphics processing items.
The personal cloud model helps context home windows of as much as 1 million tokens, whereas the general public cloud model helps as much as 256,000 tokens. Featherless additionally gives internet hosting in Europe and america and says it doesn’t hold logs.
AMD focus
A key a part of the announcement is the usage of AMD fairly than Nvidia {hardware} for native execution of GLM 5.2. Featherless says it’s the solely platform to have achieved this optimisation for the mannequin, which it argues helps prospects keep away from provide shortages and excessive procurement prices linked to Nvidia chips.
The declare displays a broader shift within the AI infrastructure market, the place demand for Nvidia processors has pushed up prices for firms looking for to deploy massive fashions at scale. By constructing round AMD, Featherless is presenting a lower-cost route for enterprises that need personal deployments with out counting on usage-based pricing from main mannequin suppliers.
Eugene Cheah, Chief Govt Officer and Co-Founding father of Featherless, outlined the corporate’s view of the economics of AI software program improvement.
“The monetary actuality of closed-source AI fashions has change into the main bottleneck in enterprise software program scalability. Spending over one million {dollars} yearly on tokens is inherently constraining engineering pace and pushing companies for his or her effectivity. The open supply framework is the best way ahead in software program improvement because it breaks vendor lock-in and gives unparalleled economics. With GLM 5.2 and AMD optimisation, we’re delivering to enterprises full technological freedom and really clear budgets,” mentioned Eugene Cheah, Chief Govt Officer and Co-Founding father of Featherless.
Mannequin efficiency
Featherless says GLM 5.2 is designed for software program engineering duties and lengthy coding periods. The mannequin makes use of a Combination-of-Consultants design with 744 billion parameters, of which 39 billion are energetic per token, and contains adjustments supposed to enhance coding and reasoning efficiency over the earlier era.
These adjustments embrace a 1 million token context window, an IndexShare mechanism designed for lengthy coding periods, and a revised multi-token prediction layer that Featherless says improves speculative decoding acceptance by about 20%. It additionally contains built-in controls that allow customers select between Excessive and Max settings for reasoning and pace trade-offs.
The mannequin is launched below the MIT licence, making it one of many extra permissively licensed massive fashions out there to companies looking for open-source alternate options. Which will matter for firms seeking to scale back dependence on closed programs whereas retaining flexibility in deployment and integration.
To assist its case for GLM 5.2 in software program improvement, Featherless cited a collection of benchmark outcomes. It mentioned Terminal-Bench 2.1 scores rose from 63.5 to 81.0 and SWE-bench Professional scores elevated from 58.4 to 62.1.
In longer and extra complicated coding exams, the reported features have been bigger. FrontierSWE rose from 30.5 to 74.4, whereas SWE-Marathon elevated from 1.0 to 13.0.
Featherless additionally 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 outcomes place GLM 5.2 among the many main open-source fashions for long-term programming duties and nearer to the efficiency of premium closed-source programs.
The launch follows Featherless’s partnership with Z.ai to distribute GLM 5.2 globally. For purchasers, which means entry to the mannequin by Featherless’s managed infrastructure fairly than having to run the structure themselves, a distinction which will enchantment to engineering groups looking for personal deployments with out constructing the underlying internet hosting stack.
Featherless says the mannequin is meant as a drop-in various for enterprise software program improvement groups that at present depend on merchandise similar to Claude Opus 4.8 and GPT-5.5. Its problem to these suppliers rests much less on uncooked benchmark scores than on whether or not a fixed-fee construction can persuade firms to commerce acquainted closed programs for an open-source mannequin working on AMD infrastructure.