Splash powers fast local Mac agents

// 1h agoINFRASTRUCTURE

Splash powers fast local Mac agents

Splash is an open-source Apple Silicon inference engine that connects local models to Codex, OpenCode, Claude Code, and compatible APIs. It uses model-specific Metal kernels, DFlash 2 speculative decoding, caching, and batching to accelerate on-device coding workloads. [GitHub](https://github.com/incoai/splash)

// ANALYSIS

Splash makes a strong case for specialized local inference, though its Apple Silicon, macOS, memory, and supported-model requirements limit its audience.

–Benchmarks report 2–5× decode gains over llama.cpp on supported Qwen models, with larger benefits under concurrent agent workloads. [GitHub releases](https://github.com/incoai/splash/releases)
–OpenAI- and Anthropic-compatible APIs make migration straightforward for existing coding-agent workflows.
–Model-specific optimization trades universal flexibility for substantially better performance and simpler tuning.
–M3-or-newer hardware and at least 36 GB of unified memory make this a premium local-development tool.
–The project’s Apache-2.0 license and integration with LM Studio broaden its potential beyond a niche CLI. [Inco AI](https://inco.ai/blog/splash/)

// TAGS

splashinferenceedge-ailocal-firstcoding-agentai-codingtool-useopen-source

DISCOVERED

1h ago

2026-10-01

PUBLISHED

1h ago

2026-10-01

RELEVANCE

9/ 10

AUTHOR

Github Awesome

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL35m ago

Tavus Griffin Passes Video Turing Test

Tavus introduced Griffin, a real-time, full-duplex video-to-video model that sees, hears, interrupts, gestures, and responds during face-to-face conversations. In Tavus’s one-minute study, 26 of 54 participants believed Griffin-Lite was a real person; the research preview is limited to selected testers.

MODEL36m ago

Tavus Griffin Passes Video Turing Test

Tavus introduced Griffin, a full-duplex Human Interaction Model that simultaneously sees, hears, speaks, and responds with humanlike timing and expression. Its Griffin-Lite research preview convinced 48% of participants they were speaking with a real person. Tavus

UPDATE38m ago

Tesana Rebuilds Game Stack Around Three.js

Tesana has rebuilt its AI game-generation platform around Three.js, replacing its Godot-based pipeline with web-native TypeScript games. The company says average generation time has fallen from roughly 60 minutes to under five.