A new release of Cohesix is trying to solve one of the messier problems in Cohesix edge AI deployments: knowing for certain which job actually ran, on which machine, and whether the result you’re looking at is real. Cohesix 1.2.0 ties job submission, monitoring, and reconciliation into a single accountable flow, so a dropped connection no longer means guessing whether a task needs to be rerun from scratch.
Summary
Key takeaways
- Cohesix 1.2.0 links job submission, monitoring, and reconciliation into one governed workflow for edge AI deployments.
- A module called “The Queen” runs on seL4 inside QEMU or on a Raspberry Pi 4, handling authority, scheduling, and recovery without touching GPU libraries.
- GPU work stays on the hardware that owns the accelerator: CUDA on NVIDIA Jetson, MLX on Apple Silicon Macs.
- Clients including SwarmUI, the command line, the Python SDK, MCP tools, and A2A agents can all follow the same governed job.
- Installation packages cover Apple Silicon Mac, JetPack Linux ARM64, and Raspberry Pi 4, with compatibility tied to specific verified release profiles.
Cohesix 1.2.0 enhances accountable edge AI workflows
Traceability sits at the heart of this release, meaning that admission, execution, and verification of every job can be reviewed after the fact instead of taken on faith. With Cohesix 1.2.0, edge AI workloads move through a single accountable flow — a job is submitted, its admission can be observed, it executes on the right host, and the output that host generates can be examined. If a connection drops mid-job, the system can reconcile the original request instead of blindly resubmitting it and risking a duplicate or a silent failure.
Unified job submission, monitoring and reconciliation
This matters because edge deployments are notoriously unreliable at the network layer — Wi-Fi drops, devices reboot, connections time out. Rather than treating every interruption as a failure requiring a retry, Cohesix keeps the job’s identity intact and reconciles it against what actually happened on the hardware. That distinction between “the request was accepted” and “the request was completed” runs through the entire release.
The Queen module for authority, scheduling, and recovery
At the center of this governance layer sits a module the release notes call “The Queen.” It runs inside seL4, a security-focused microkernel, within either a QEMU virtual machine or directly on a Raspberry Pi 4. The Queen’s job is narrow but important: it applies authority, scheduling, and recovery rules across the system. Critically, it does not load GPU libraries into the VM or the Pi itself — that work is deliberately kept elsewhere.
GPU acceleration and workflow management
GPU acceleration in Cohesix 1.2.0 is handled entirely on the hardware that owns the accelerator, not inside the virtualized control layer. That separation is what allows GPU job scheduling to stay predictable even as the orchestration layer changes.
GPU work stays on the host owning the accelerator
Concretely, CUDA acceleration runs on the supported NVIDIA Jetson reference hardware, while MLX runs on a supported Apple Silicon Mac using Apple’s own Metal runtime. The Queen never touches these GPU libraries directly; it coordinates scheduling and recovery, while the actual compute stays local to the device that has the accelerator.
Support for CUDA on NVIDIA Jetson and MLX on Apple Silicon Macs
On the Jetson side, the supported CUDA path covers inference and adapter work. On Apple hardware, the local MLX path taps into Metal directly. Provider selection is explicit in Cohesix, meaning a job cannot quietly shift from MLX to CUDA, or fall back to CPU, without that change being visible.
Job admission and native GPU provider result verification
Perhaps the most consequential design choice in this release is how Cohesix separates job admission from job completion. An admission acknowledgement simply means the Queen accepted the request — it does not mean CUDA or MLX has actually finished the work. Cohesix retains the original job identity and checks the native provider’s result independently. A lost reply, an uncertain outcome, or a failed candidate stays visible in the system rather than being silently treated as a success. For anyone running production inference at the edge, that distinction between acceptance and verified completion is exactly the kind of detail that prevents a bad result from slipping through unnoticed.
Client ecosystem and model lifecycle management
Cohesix 1.2.0 doesn’t require switching tools — it’s built to work with whatever client an operator already uses, while still enforcing the same governance rules underneath.
Compatible clients: SwarmUI, command line, Python SDK, MCP tools, and A2A agents
SwarmUI, the Cohesix command line and Python clients, MCP tools, and A2A agents can all follow the same governed job workflow. On the QEMU and Raspberry Pi release profiles, MCP and A2A are enabled by default in the host gateway, though operators can disable either protocol in their own deployment profile if they don’t need it.
PEFT workflow for managing training, evaluation, canary releases, promotion, and rollback
Model updates get their own dedicated path. The PEFT workflow connects training or import directly to evaluation, canary release, promotion, and rollback — giving operators a structured way to manage PEFT model management without ad hoc scripting. Notably, a rejected candidate model stays rejected even if an earlier serving version is later restored, which closes off a scenario where a bad model could quietly resurface after a rollback.
Installation and compatibility considerations
Cohesix installation in version 1.2.0 spans three hardware targets plus a standalone Python package, giving teams a fairly direct path from download to a running deployment.
Available installation packages for Apple Silicon Mac, JetPack Linux ARM64, Raspberry Pi 4, and Python clients
For Apple Silicon Macs, there’s a portable tar.gz archive or a signed and notarized .pkg installer that places SwarmUI in Applications. For JetPack Linux ARM64, operators can choose a portable tar.gz build or Debian packages — a controller package plus an optional SwarmUI package — with the Debian publisher signing the package manifest to bind both hashes together. Raspberry Pi 4 users get an SD card image with its own checksum for flashing. The Python SDK and the NeMo Agent Toolkit client kit ship inside the host archives and are also published separately on PyPI as cohesix==1.2.0. Both the Mac and Linux installers bundle the Cohesix command line tools and SwarmUI, and QEMU is optional — the setup process checks for an installed binary and offers a compatible pinned build if the system package doesn’t fit. No installer downloads a model or copies credentials during setup.
Compatibility requires qualification beyond verified release profiles
The release is explicit about where its guarantees stop. On ARM64, the Linux installer uses JetPack 7.2.1 / L4T 39.2.1 as its reference configuration, while Apple Silicon serves as the reference for Mac. For Raspberry Pi 4, qualification depends on wired GENET networking instead of Wi-Fi. Outside of those specific profiles — other JetPack versions, different Linux distributions, or Pi setups running over Wi-Fi — compatibility isn’t assumed, and matching CPU architecture alone doesn’t count as verification. In practice, that means teams running anything outside these exact reference configurations should expect to qualify their own setup before relying on it in production.
FAQ
What environments does the Cohesix Queen module run on?
The Queen runs in seL4 on QEMU or Raspberry Pi 4 environments.
How does Cohesix ensure the accuracy of GPU job results?
Cohesix distinguishes between job admission by the Queen and native GPU provider completion by separately verifying results.
Which hardware platforms are supported for GPU acceleration in Cohesix 1.2.0?
Cohesix supports CUDA on NVIDIA Jetson devices and Apple Metal (MLX) on Apple Silicon Macs.
How is model lifecycle managed within Cohesix 1.2.0?
The PEFT workflow manages the training, evaluation, canary release, promotion, and rollback of models deliberately.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.




