LocalAI v4.10.0 Released: Fleet Dashboard & M5 Support

September 18, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

LocalAI version v4.10.0 has been released. LocalAI is an open-source AI engine (MIT license) written in Go that runs generative AI models such as LLMs, images, audio, and video locally on any hardware without requiring a GPU.

This release introduces a fleet operation dashboard for monitoring and managing the state of an entire cluster, credential file management for private model sources and registries, a benchmark command for measuring performance from the CLI, and per-model environment variable settings. It also includes fixes for CVE vulnerabilities and a startup crash on Apple M5.

Key Changes

New Features

• Fleet Operation Dashboard: The previous flat node list has been revamped into a dashboard that provides real-time visibility into cluster-wide health bands, VRAM/RAM/CPU/disk capacity gauges, and replica placement of running models. It also features batched operations with concurrency limits (Drain/Resume/Remove) and a detailed inspector.
• Credentials File for Private Sources: Using the --credentials-file option or the LOCALAI_CREDENTIALS_FILE environment variable, a single credentials.yaml can now centrally manage authentication for OCI registries, galleries, HTTP downloads, and GitHub repositories.
• CLI Benchmark Command: The local-ai benchmark MODEL command has been added, allowing direct measurement of latency and throughput (completion tokens/sec) for configured text models straight from the CLI.
• Per-Model Environment Variables: By defining an env: map in model configuration files, environment variables such as CUDA_VISIBLE_DEVICES and VLLM_CACHE_DIR can be injected into the process on a per-model basis.
• Mid-Stream System Message Control: Added the template.system_messages_after_first setting. To avoid errors when system messages are added mid-conversation in models like Qwen3.x, it can either merge them into the first system message ("merge") or convert them into user utterances ("user").
• Context Size Exposure: The /v1/models/capabilities endpoint now returns ContextSize, allowing clients to directly check the effective context window of each model.

Performance Improvements

• Load-Time LoRA Fusion in vllm-cpp: The vllm-cpp video backend now consumes lora_adapters and lora_scales settings to fuse LoRA deltas into DiT weights during engine loading (supported for LTX2.5 and MiniMax-H3).

Bug Fixes and Security

• Apple M5 Startup SIGSEGV Fix: Updated go-m1cpu to v0.2.2, fixing a segmentation fault that occurred at startup on Apple M5 hardware.
• Security Vulnerability Fixes: Patched a total of 4 CVE vulnerabilities in ip-address (CVE-2026-69192), containerd (CVE-2026-46680, CVE-2026-53488), and react-router (GHSA-qwww-vcr4-c8h2).

Supported Models and Hardware

• New Model Galleries: 140 gallery entries have been added, including Qwopus 3.6, OpenResearcher, LFM2.5 DSpark, Qwen3.8 small distills / Flash Next, UI-Mate 27B, Homura 30B, Ornith 1.5 35B/9B, Granite 4.2, Tiel-Coder 35B, Thomson 1.0 Small, PhoneLLM, WeMM embedding, EXL3 vllm.cpp, Gemma 4 E4B, and Orukeet for NeMo speech.
• Hardware Support: Includes startup fixes for Apple M5 and ROCm build support for stablediffusion-ggml.

How to Get It

For detailed update instructions, please refer to the LocalAI v4.10.0 Release Page.

Releases Since Our Last Article

Compiled by Local Model Watch from the project’s GitHub releases: the versions between this release and the last one we covered, which did not get separate articles. Full history: release tracker.

Sources

• https://github.com/mudler/LocalAI/releases/tag/v4.10.0