{"slug": "cohesix-find-out-what-happened-to-a-local-ai-job", "title": "Cohesix – find out what happened to a local AI job", "summary": "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.", "body_md": "# Cohesix 1.2.0\n\nCohesix 1.2.0 brings edge AI work into one accountable flow. Submit a job,\n\nwatch the Queen admit it, run it on the right host, and inspect the result that\n\nhost actually produced. If a connection drops after submission, Cohesix can\n\nreconcile the original job instead of guessing whether it should run again.\n\nThe Queen runs in seL4 on QEMU or Raspberry Pi 4. GPU work stays on the host\n\nthat owns the accelerator: CUDA on the supported NVIDIA Jetson reference and\n\nMLX on a supported Apple Silicon Mac. The Queen applies authority, scheduling\n\nand recovery rules; it does not load GPU libraries into the VM or Pi.\n\n## What you can do\n\n- **Use the client you already have.** SwarmUI, Cohesix command line and\n Python clients, MCP tools, and A2A agents can follow the same governed job.\n The QEMU and Pi release profiles enable MCP and A2A in the host gateway by\ndefault. Operators can disable either protocol in their deployment profile.\n- **Know which result is real.** An admission acknowledgement means the Queen\n accepted a request; it does not mean CUDA or MLX finished. Cohesix retains\n the original job identity and checks the native provider result separately.\n A lost reply, uncertain outcome or failed candidate stays visible rather\nthan being silently replayed as success.\n- **Run useful GPU workflows.** The supported CUDA path covers inference and\n adapter work on the Jetson reference. The local MLX path uses Apple's Metal\n runtime on Apple Silicon. Provider selection is explicit, so a job does not\nquietly move from MLX to CUDA or CPU.\n- **Manage model changes deliberately.** The PEFT workflow connects training\n or import to evaluation, canary, promotion and rollback. A rejected\n candidate remains rejected even if the earlier serving version is restored.\n The pinned NeMo Agent Toolkit kit provides MCP and A2A client examples while\nkeeping client dependencies separate from the CUDA provider environment.\n\n## Pick a download\n\n| Where you run it | Release files | \n|---|---|\n| Apple Silicon Mac | `Cohesix-1.2.0-MacOS.tar.gz` for a portable setup, or the signed and notarized`Cohesix-1.2.0-MacOS.pkg` for a normal Mac installation with SwarmUI in Applications. | \n| JetPack Linux ARM64 | `Cohesix-1.2.0-linux.tar.gz` for a portable setup, or`cohesix-controller_1.2.0_arm64.deb` plus the optional`cohesix-swarmui_1.2.0_arm64.deb` for a system installation. The Debian publisher signs the package manifest, which binds both package hashes. | \n| Raspberry Pi 4 | `Cohesix-1.2.0-Pi4.tar.gz` contains`image/cohesix-pi4-sd.img` and its checksum for flashing an SD card. | \n| 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 as`cohesix==1.2.0` . | \n\nThe Mac and Linux host installers include the Cohesix command line tools and\n\nSwarmUI. QEMU is optional: setup checks the installed binary and offers a\n\ncompatible pinned build when the system package is unsuitable. No installer\n\ndownloads a model or copies your credentials.\n\nStart with the [quickstart](https://github.com/lukeb-aidev/cohesix/blob/v1.2.0/docs/QUICKSTART.md), then use the\n\n[use cases](https://github.com/lukeb-aidev/cohesix/blob/v1.2.0/docs/USE_CASES.md) to choose a client and provider. Verify a\n\ndownload against the release checksums and the package signature before\n\ninstalling it. The Debian publisher key fingerprint is\n\n`6840 71A2 AF49 8008 930C F6AE EBB8 ECC4 9397 F8C2`.\n\n## Supported release profiles\n\nThe Linux installer reference is JetPack 7.2.1 / L4T 39.2.1 on ARM64. The Mac\n\nreference is Apple Silicon. Raspberry Pi 4 release qualification uses wired\n\nGENET networking. Other JetPack versions, Linux distributions and Pi Wi-Fi\n\nsetups require their own qualification; the matching CPU architecture alone\n\ndoes not establish compatibility. See [host tools](https://github.com/lukeb-aidev/cohesix/blob/v1.2.0/docs/HOST_TOOLS.md) and\n\n[GPU nodes](https://github.com/lukeb-aidev/cohesix/blob/v1.2.0/docs/GPU_NODES.md) for the selected host and provider contracts.\n\nFor the exact release artifact hashes, installation checks and target evidence,\n\nsee the [release evidence checklist](https://github.com/lukeb-aidev/cohesix/blob/v1.2.0/docs/audit/checklists/RELEASE_EVIDENCE_CHECKLIST.md).", "url": "https://wpnews.pro/news/cohesix-find-out-what-happened-to-a-local-ai-job", "canonical_source": "https://github.com/lukeb-aidev/cohesix/releases/tag/v1.2.0", "published_at": "2026-09-28 12:05:22+00:00", "updated_at": "2026-09-28 12:17:59.517915+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-agents", "agent-protocols", "mlops", "developer-tools"], "entities": ["Cohesix", "seL4", "QEMU", "Raspberry Pi 4", "NVIDIA Jetson", "Apple Silicon", "SwarmUI", "NeMo Agent Toolkit"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/cohesix-find-out-what-happened-to-a-local-ai-job", "markdown": "https://wpnews.pro/news/cohesix-find-out-what-happened-to-a-local-ai-job.md", "text": "https://wpnews.pro/news/cohesix-find-out-what-happened-to-a-local-ai-job.txt", "jsonld": "https://wpnews.pro/news/cohesix-find-out-what-happened-to-a-local-ai-job.jsonld"}}