Coding Horizon

Local AI Just Got Voice, Vision, Video And Routing

Everything this video puts on screen as a name, a figure, a version or a date, chased back to a primary source. Retrieved 11 August 2026.

What LocalAI says it is

“LocalAI is the open-source AI engine. Run any model - LLMs, vision, voice, image, video - on any hardware. No GPU required.”

The project’s own repository description and the first line of its README.

“Make AI run on every machine.” The tagline on the project site, which expands it to “Text, voice, vision, images, video, 3D and agents, from one open runtime. It works on the laptop you already own, and scales to a room full of GPUs when you have one.” The site organises the feature set into six pillars: Reason, Listen, Speak, See, Create, Act. It also states that “Every feature ships a CPU path first”, which is the basis of the no GPU claim, and that the server speaks the OpenAI, Anthropic, Ollama and ElevenLabs APIs.

Licence: MIT. 48,391 stars, 4,348 forks. First commit March 2023; the repository was created on 18 March 2023 and was last pushed to on 11 August 2026, the day these figures were read.

60+ backends, including llama.cpp, vLLM, SGLang, transformers, whisper.cpp, diffusers, MLX and MLX-VLM, alongside native C/C++/GGML engines written by the project itself.

The tools that made local language models accessible

llama.cpp, Ollama and LM Studio are named as the first wave that put local models on ordinary machines. LocalAI itself lists llama.cpp among its backends, and its API surface includes a drop-in Ollama-compatible API.

The release timeline

Nine minor releases between March and August 2026, each with its own write-up on the project blog:

Version Date What it added
4.0 14 Mar 2026 Agent orchestration in the core, React interface
4.1 2 Apr 2026 Distributed cluster mode, OIDC per user API keys
4.2 11 May 2026 Voice recognition, face recognition, speaker diarization
4.3 24 May 2026 Prompt cache on by default, usage per API key, Distributed v3
4.4 10 Jun 2026 Prefix cache aware routing, video input and video generation
4.5 23 Jun 2026 Prefix caching by default, speaker aware conversations
4.6 4 Jul 2026  
4.7 14 Jul 2026  
4.8 5 Aug 2026 New inference engine, terminal agent CLI, 3D generation

The most recent tag at the time of writing is v4.8.2, 7 August 2026.

Voice, face and diarization

LocalAI 4.2.0, 11 May 2026. Adds /v1/voice/* endpoints for speaker verification, identification, embedding extraction and analysis; a face recognition pipeline covering 1:1 verification, 1:N identification, detection, analysis and embeddings, with liveness detection to reject spoofed photos and videos; and a /v1/audio/diarization endpoint that answers “who spoke when”, backed by sherpa-onnx and vibevoice.cpp. The same release brings video generation to the stable-diffusion.ggml backend, including image to video and first-last-frame generation, and eleven new backends in total.

Prompt caching

LocalAI 4.3.0, 24 May 2026. The llama.cpp server side prompt cache is enabled automatically rather than left off. The release states the effect directly: “Repeated system prompts go from 5-8 min to seconds.”

LocalAI 4.5.0, 23 June 2026 extends this to prefix caching on by default, so system prompts, RAG context, agent scaffolds and multi-turn chat are not recomputed on every request, which the release describes as a time to first token and throughput win for shared prefix workloads.

Usage attribution

LocalAI 4.3.0, 24 May 2026. Traffic is attributed per API key and per user, surfaced in the interface, and revoked keys stay readable in the history rather than taking their traffic with them. Per user API key management arrived earlier, in 4.1.0, 2 April 2026, via OIDC.

Distributed routing

LocalAI 4.1.0, 2 April 2026 introduced distributed cluster mode. 4.3.0, 24 May 2026 shipped Distributed v3 with per request replica routing, cached health probes and async per node installation with streaming progress. 4.4.0, 10 June 2026 turned on prefix cache aware routing by default, which biases a request toward the replica that already holds the relevant KV and prefix cache.

LocalAI 4.5.0, 23 June 2026 scales parallel serving slots to available VRAM, up to eight on systems with 32GB or more, so several users share one engine without multiplying KV memory.

Beyond text

LocalAI 4.4.0, 10 June 2026 adds video input understanding in llama-cpp via mtmd and video generation via LTX-2 in stablediffusion-ggml, plus the native rfdetr-cpp detection backend with 32 prebuilt GGUFs. 4.5.0 adds monocular metric depth via a depth-anything backend, sound event tagging across 527 AudioSet classes, an espeak free ONNX text to speech backend, and speaker aware realtime conversations that surface the recognised speaker to both the client and the model.

Not checked