How the Ethereum Foundation dAI team secures open-weight models using Featherless and Averta. Summary. Featherless and Averta work together to make open model inference enterprise-ready. Featherless provides serverless access to a large and expanding library of open-weight models, while Averta adds the runtime security and governance layer needed to protect AI systems in production. Through the live webhook integration, Featherless can apply Averta Guardrails and policy controls across the model access it provides. The Ethereum Foundation dAI team's use of Featherless and Averta together for research purposes provides clear validation of the partnership. The challenge: open models need enterprise-grade security. For teams building AI products, model access has become a practical infrastructure decision. Closed model providers offer simple APIs, but limit teams to a smaller set of models. GPU hosting platforms offer greater flexibility, but often require teams to manage infrastructure and operational costs themselves. Featherless addresses this gap as a serverless provider with specialized model loading and GPU orchestration capabilities that allow it to keep an exceptionally large catalogue of models online. It combines a broad model range and variety with serverless pricing, giving teams low-cost, easy access to a continually expanding library of open-weight models without requiring them to operate their own inference infrastructure. As Featherless lowers the cost and operational burden of accessing open-weight models, the next challenge is securing the AI systems built on top at runtime. Native guardrails and system prompts are insufficient on their own because they were not designed to govern the full AI execution path. In production, AI systems may handle sensitive data, call tools, and trigger high-risk actions. Each interaction needs enforceable policy, monitoring, governance, and auditability. This is the problem Averta solves: protecting AI systems across the execution path, where classification, policy enforcement, access control, and auditability work together as a single security layer. Without that runtime layer, AI applications remain exposed to predictable failure modes: prompt injection, unauthorised tool use, sensitive data leakage, policy bypasses, and other real-world attacks. These risks become more serious when agents are embedded in workflows, connected to tools, or allowed to trigger actions on behalf of users. Governance expectations are also rising, with standards such as ISO/IEC 42001 and regulations such as the EU AI Act reinforcing the need for AI systems to be risk-managed, observable, auditable, and resilient in production. Enterprise readiness requires both sides of the equation: scalable access to models and a security layer that protects the full AI execution path, from input analysis and policy enforcement to runtime security and auditability. The collaboration: Featherless and Averta. The Featherless and Averta integration connects serverless open model inference with Averta's runtime guardrails and policy controls. Through the live webhook integration, Featherless now routes AI activity through Averta's runtime checkpoints, applying policy decisions around prompts, tool calls, tool results, and outputs across the AI execution path. Featherless manages the policies applied through this integration, maintaining a consistent security layer across the model access it provides. This allows teams to use Featherless' open-weight model access while applying Averta Guardrails to the production workflows, agents, and applications connected to those models. The partnership between Averta and Featherless exists to secure the open-weight models Featherless serves. Security is moving up the priority list for organizations integrating AI, and it is easier to build in at the inference layer than to bolt on afterward. Supporting Ethereum Foundation dAI team research. The Ethereum Foundation dAI team is using Featherless and Averta together for research purposes, providing a clear validation point for the partnership. Featherless delivers serverless access to open-weight models, while Averta Guardrails applies runtime security and policy enforcement to the AI workflows using them. For the EF dAI team, this means security decisions can be made across the AI execution path: before requests reach the model, before tools are exposed or called, when tool results return, and before final outputs reach users. Averta Guardrails helps enforce policy, monitor behaviour, and provide auditability at each of these points. The deployment shows how Featherless and Averta work together in a high-trust environment: Featherless provides the model access layer, and Averta provides the security and governance layer needed to evaluate agentic AI workflows safely and reliably. Outcomes and impact. The webhook integration between Featherless and Averta is live, enabling AI activity from Featherless to be routed into Averta Guardrails and policy controls. The Ethereum Foundation dAI team is using Featherless and Averta together for research activities, demonstrating how open model inference can be paired with runtime AI security in a high-trust environment. Averta Guardrails is built for the AI execution path, where classification has to be fast, precise, and consistent. The classification engine reaches 98.8% precision on adversarial and benign traffic, evaluated against held-out attack corpora with less than 40ms p99 latency. It provides 100% action coverage across prompts, tool calls, and outputs, with zero unclassified executions silently allowed. The security impact is visible in attack success rate testing across frontier models. Across the models tested, Averta Guardrails reduced average attack success rates from 73% to 4% for prompt attacks, 58% to 6% for unauthorised actions, 41% to 2% for data exfiltration, and 67% to 5% for policy violations. Averta's broader platform extends this security and governance layer across the AI stack. MCP Gateway helps organisations manage enterprise agents, agent identity and ownership, agent activity, and agent tool access across popular integrations such as Linear, GitHub, and Notion, across the entire workforce. Averta RED adds expert-led and automated AI red teaming for launch reviews, regression checks, and ongoing validation, simulating the prompt injection, tool abuse, and data exfiltration paths agents are likely to face in production. Together, these results show the role Averta Guardrails plays alongside Featherless' serverless inference layer: adding runtime protection, policy enforcement, and auditability to open model inference in production. For organisations with broader agent security needs, Averta also offers MCP Gateway for agent and tool governance and Averta RED for continuous testing and validation.

Featherless.ai partners with Z.ai to provide global access to new open-source AI model, GLM 5.2. Featherless.ai has partnered with Z.ai to deliver worldwide access to the recently launched GLM 5.2 AI model. With this collaboration, organisations can seamlessly deploy and run the model through Featherless.ai, without encountering any hindrances related to the infrastructure necessary for hosting large-scale open-source architectures. Featherless.ai is a Day Zero launch partner for GLM-5.2. The model is available through the OpenAI-compatible API, in FP8 with up to a 256K context window on public cloud and up to 1M on private cloud deployments. Featherless.ai hosts the model directly, which means no need for GPU provisioning, with no logs and with hosting in the EU and US. The launch takes place at a time when regulation and competition in AI are fierce. The U.S. government has just imposed export controls on the highest-level frontier model of Anthropic, the Fable 5. As an open-source alternative to closed-source technology, GLM 5.2 reduces the gap in competition. It is projected that GLM 5.2's initial evaluation will place it at a comparable level with that of Claude Opus 4.7-4.8. Eugene Cheah, CEO and co-founder of Featherless.ai, said: "Any attempt to enforce tight export controls or closed architectures as a way to keep the frontier features in check will only result in the development of an alternative. In the situation where an engineering team or company feels that the access might be revoked suddenly, there is an incentive to switch to technologies that they can operate and change entirely independently of anyone else. That's precisely why open source technology evolves so quickly." The structural abilities of GLM 5.2 are significantly improved and include a Mixture-of-Experts design with 744 billion parameters that use 39 billion parameters per token. The GLM-5.2 structure is similar in physical size to its predecessor, GLM-5.1. Performance is improved on a generation-by-generation basis thanks to careful coding agent training and an optimised 1 million token context window. To ensure quality during long coding agent sessions, the architecture includes IndexShare, which allows using a common lightweight indexer in each fourth sparse attention layer. Integration helps save compute costs by 2.9 times at high context lengths. Also, an updated multi-token prediction layer improves speculative decoding acceptance by about 20%, whereas built-in thinking effort control lets choosing between High and Max levels for reasoning vs. real-time trade-off. The model is released under the permissive MIT license and available through an OpenAI-compatible API. The architecture improvements yield tangible performance gains under several rigorous evaluation criteria. In long-term programming tasks such as FrontierSWE, PostTrainBench, and SWE-Marathon, GLM-5.2 becomes the top-ranked open-source model in the industry. It is the industry's first true open-source, drop-in replacement for enterprise software development capable of competing directly with closed systems like Claude Opus 4.8 and GPT-5.5 - something major engineering leaders and tech executives (Vercel, Tesla) have also taken notice of. Featherless.ai is introducing an exclusive private cloud deployment that pairs GLM-5.2 with AMD infrastructure to create a secure, fully compliant and highly cost-efficient software engineering environment. As the only platform to have optimised GLM-5.2 to run natively on AMD hardware, Featherless.ai enables organisations to completely bypass ongoing NVIDIA chip shortages and inflated procurement costs while maximising AMD's substantial compute-per-dollar advantage. As empirical testing reveals, Terminal-Bench 2.1 performance scores went from 63.5 to 81.0, and SWE-bench Pro scores rose from 58.4 to 62.1. The scores in complex development assessments were even more dramatic, with FrontierSWE rising from 30.5 to 74.4 and SWE-Marathon going from 1.0 to 13.0. In pure logical reasoning tasks, the scores went up as well, with AIME 2026 going from 95.3 to 99.2 and GPQA-Diamond scores rising from 86.2 to 91.2.