Docker Desktop 4.83.0 shipped on July 20, and the update that matters for developers building AI features locally is not the security patches — it is the inference inspector now baked into Docker Model Runner. It shows you the exact HTTP payload your application sends to a local model and the exact response that comes back. No proxy. No logging middleware. No digging through llama.cpp debug output. If you have ever spent an hour debugging a RAG pipeline only to discover your embedding call was silently truncating input, this is the fix.
Docker Model Runner in 90 Seconds
Docker Model Runner (DMR) is a local LLM inference server built into Docker Desktop since version 4.40. It runs models as native processes on host hardware — Metal on Mac, CUDA on Linux — using llama.cpp under the hood, and exposes an OpenAI-compatible API at localhost:12434. Models are packaged as OCI artifacts and pulled the same way as container images:
docker model pull ai/llama3.2
docker model run ai/llama3.2 "Explain this stack trace"Because the API is OpenAI-compatible, you write your application against the familiar /v1/chat/completions endpoint, point it at localhost during development, and swap the base URL to the real OpenAI API in production without touching the rest of your code.
The Inference Inspector: See What Your App Is Actually Sending
The inference inspector is accessible from the Model Runner section inside Docker Desktop. When active, it captures every inference request routed through DMR and surfaces the full HTTP payload: prompt content, system messages, headers, and the raw model response. This is the observability layer that was missing from local LLM development.
The practical payoff is immediate. Prompt formatting bugs are the single biggest time sink in local AI development — a missing newline between system and user messages, a role field set to the wrong value, a context window being exceeded without a clear error. The inspector makes all of that visible in real time without switching tools.
To enable it: navigate to Settings, select Features in Development, enable Model Runner, and apply the restart. The inspector activates automatically once Model Runner is running. Fire a request from your application and the payload appears in the panel.
Run Embedding and Inference Models at the Same Time
4.83 also adds resource-aware concurrent model loading. You can now run an embedding model alongside a chat model without the second load starving the first:
# Load both models in background
docker model run --detach ai/nomic-embed-text
docker model run --detach ai/llama3.2
# Embed your documents
curl localhost:12434/engines/v1/embeddings -H "Content-Type: application/json" -d "{"model": "ai/nomic-embed-text", "input": "your document text"}"
# Query with the chat model
curl localhost:12434/engines/v1/chat/completions -H "Content-Type: application/json" -d "{"model": "ai/llama3.2", "messages": [{"role": "user", "content": "your query"}]}"Docker Desktop now checks available resources before loading a second model and surfaces a clear error instead of silently failing or causing an OOM event. For teams building RAG pipelines with Docker Model Runner, this removes the last practical reason to reach for an external embedding service during development.
Registry Mirrors, Two CVEs, and One Breaking Change
Docker Model Runner v1.2.6 — bundled in this release — adds registry mirror support. You configure a private mirror for model distribution the same way you configure Docker Hub mirrors today. This unlocks air-gapped deployments, lets platform teams control which model versions enter the environment, and speeds up pulls when the mirror is co-located with your build infrastructure.
Two security fixes shipped in 4.83 that warrant immediate attention. CVE-2026-8936 is a VM panic from unbounded recursion in the grpcfuse kernel module, triggered by containers creating deeply nested directories on bind-mounted host folders. CVE-2026-31431 is the more serious one: an unprivileged container user could gain root inside the container VM via a controlled write into the host VM page cache. If you run any untrusted container workloads, update today. Full details are in the Docker Desktop release notes.
One breaking change to note: docker sbom is deprecated. Switch to docker scout sbom before it is removed in a future release.
Docker Model Runner vs Ollama: The Honest Take
If you are starting a new project and want the fastest path to a running local LLM, Ollama remains the right default. It has a simpler install story, a large integration ecosystem, and excellent Apple Silicon performance. Ollama just raised million on the strength of local AI adoption, which is a strong signal that the space is real.
Docker Model Runner wins when you are already deep in Docker workflows. Models follow the same OCI packaging as your containers, you pull them with docker model pull, and they slot naturally into your existing CI/CD setup. Both tools can coexist — reach for Ollama to explore, reach for DMR when you need models in the same pipeline as your containers. The inference inspector, concurrent model loading, and the Docker Model Runner general availability announcement make 4.83 the most compelling version of DMR yet.
Update to 4.83.0 from the Docker Desktop application menu or via your package manager. The CVE-2026-31431 privilege escalation fix makes this a mandatory update regardless of whether you use Model Runner.
