cd /news/ai-infrastructure/cohesix-find-out-what-happened-to-a-… · home › topics › ai-infrastructure › article
[ARTICLE · art-140963] src=github.com ↗ pub= topic=ai-infrastructure verified=true sentiment=· neutral

Cohesix – find out what happened to a local AI job

Cohesix released version 1.2.0, an edge AI job orchestration system that runs its Queen controller in seL4 on QEMU or Raspberry Pi 4 while keeping GPU work on the host that owns the accelerator, with CUDA on the supported NVIDIA Jetson reference and MLX on Apple Silicon Macs. The release adds MCP and A2A protocol support in the host gateway by default, a PEFT workflow connecting training or import to evaluation, canary, promotion and rollback, and separate verification of native provider results so a lost reply or failed candidate is not silently replayed as success. Cohesix 1.2.0 ships as signed and notarized macOS packages, Debian arm64 packages signed by the publisher key fingerprint 6840 71A2 AF49 8008 930C F6AE EBB8 ECC4 9397 F8C2, a Raspberry Pi 4 SD image, and a Python SDK published on PyPI as cohesix==1.2.0.

read3 min views5 publishedSep 28, 2026
Cohesix – find out what happened to a local AI job
Image: Michielbdejong (auto-discovered)

Cohesix 1.2.0 brings edge AI work into one accountable flow. Submit a job,

watch the Queen admit it, run it on the right host, and inspect the result that

host actually produced. If a connection drops after submission, Cohesix can

reconcile the original job instead of guessing whether it should run again.

The Queen runs in seL4 on QEMU or Raspberry Pi 4. GPU work stays on the host

that owns the accelerator: CUDA on the supported NVIDIA Jetson reference and

MLX on a supported Apple Silicon Mac. The Queen applies authority, scheduling

and recovery rules; it does not load GPU libraries into the VM or Pi.

What you can do #

  • Use the client you already have. SwarmUI, Cohesix command line and Python clients, MCP tools, and A2A agents can follow the same governed job. The QEMU and Pi release profiles enable MCP and A2A in the host gateway by default. Operators can disable either protocol in their deployment profile.
  • Know which result is real. An admission acknowledgement means the Queen accepted a request; it does not mean CUDA or MLX finished. Cohesix retains the original job identity and checks the native provider result separately. A lost reply, uncertain outcome or failed candidate stays visible rather than being silently replayed as success.
  • Run useful GPU workflows. The supported CUDA path covers inference and adapter work on the Jetson reference. The local MLX path uses Apple's Metal runtime on Apple Silicon. Provider selection is explicit, so a job does not quietly move from MLX to CUDA or CPU.
  • Manage model changes deliberately. The PEFT workflow connects training or import to evaluation, canary, promotion and rollback. A rejected candidate remains rejected even if the earlier serving version is restored. The pinned NeMo Agent Toolkit kit provides MCP and A2A client examples while keeping client dependencies separate from the CUDA provider environment.

Pick a download #

Where you run it Release files
Apple Silicon Mac Cohesix-1.2.0-MacOS.tar.gz for a portable setup, or the signed and notarizedCohesix-1.2.0-MacOS.pkg for a normal Mac installation with SwarmUI in Applications.
JetPack Linux ARM64 Cohesix-1.2.0-linux.tar.gz for a portable setup, orcohesix-controller_1.2.0_arm64.deb plus the optionalcohesix-swarmui_1.2.0_arm64.deb for a system installation. The Debian publisher signs the package manifest, which binds both package hashes.
Raspberry Pi 4 Cohesix-1.2.0-Pi4.tar.gz containsimage/cohesix-pi4-sd.img and its checksum for flashing an SD card.
Python clients The Cohesix SDK and NeMo client kit are included in the host archives at version 1.2.0 . The SDK is also published on PyPI ascohesix==1.2.0 .

The Mac and Linux host installers include the Cohesix command line tools and

SwarmUI. QEMU is optional: setup checks the installed binary and offers a

compatible pinned build when the system package is unsuitable. No installer

downloads a model or copies your credentials.

Start with the quickstart, then use the use cases to choose a client and provider. Verify a

download against the release checksums and the package signature before

installing it. The Debian publisher key fingerprint is

6840 71A2 AF49 8008 930C F6AE EBB8 ECC4 9397 F8C2.

Supported release profiles #

The Linux installer reference is JetPack 7.2.1 / L4T 39.2.1 on ARM64. The Mac

reference is Apple Silicon. Raspberry Pi 4 release qualification uses wired

GENET networking. Other JetPack versions, Linux distributions and Pi Wi-Fi

setups require their own qualification; the matching CPU architecture alone

does not establish compatibility. See host tools and

GPU nodes for the selected host and provider contracts.

For the exact release artifact hashes, installation checks and target evidence, see the release evidence checklist.

── more in #ai-infrastructure 4 stories · sorted by recency
── more on @cohesix 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/cohesix-find-out-wha…] indexed:0 read:3min 2026-09-28 · —